Loading...
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] __cpuinitdata = {
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
61};
62
63int __cpuinit 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 __cpuinit 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 if (node != NUMA_NO_NODE)
101 set_cpu_numa_node(cpu, node);
102}
103
104void __cpuinit numa_clear_node(int cpu)
105{
106 numa_set_node(cpu, NUMA_NO_NODE);
107}
108
109/*
110 * Allocate node_to_cpumask_map based on number of available nodes
111 * Requires node_possible_map to be valid.
112 *
113 * Note: node_to_cpumask() is not valid until after this is done.
114 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
115 */
116void __init setup_node_to_cpumask_map(void)
117{
118 unsigned int node, num = 0;
119
120 /* setup nr_node_ids if not done yet */
121 if (nr_node_ids == MAX_NUMNODES) {
122 for_each_node_mask(node, node_possible_map)
123 num = node;
124 nr_node_ids = num + 1;
125 }
126
127 /* allocate the map */
128 for (node = 0; node < nr_node_ids; node++)
129 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
130
131 /* cpumask_of_node() will now work */
132 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
133}
134
135static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
136 struct numa_meminfo *mi)
137{
138 /* ignore zero length blks */
139 if (start == end)
140 return 0;
141
142 /* whine about and ignore invalid blks */
143 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
144 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
145 nid, start, end);
146 return 0;
147 }
148
149 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
150 pr_err("NUMA: too many memblk ranges\n");
151 return -EINVAL;
152 }
153
154 mi->blk[mi->nr_blks].start = start;
155 mi->blk[mi->nr_blks].end = end;
156 mi->blk[mi->nr_blks].nid = nid;
157 mi->nr_blks++;
158 return 0;
159}
160
161/**
162 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
163 * @idx: Index of memblk to remove
164 * @mi: numa_meminfo to remove memblk from
165 *
166 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
167 * decrementing @mi->nr_blks.
168 */
169void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
170{
171 mi->nr_blks--;
172 memmove(&mi->blk[idx], &mi->blk[idx + 1],
173 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
174}
175
176/**
177 * numa_add_memblk - Add one numa_memblk to numa_meminfo
178 * @nid: NUMA node ID of the new memblk
179 * @start: Start address of the new memblk
180 * @end: End address of the new memblk
181 *
182 * Add a new memblk to the default numa_meminfo.
183 *
184 * RETURNS:
185 * 0 on success, -errno on failure.
186 */
187int __init numa_add_memblk(int nid, u64 start, u64 end)
188{
189 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
190}
191
192/* Initialize NODE_DATA for a node on the local memory */
193static void __init setup_node_data(int nid, u64 start, u64 end)
194{
195 const u64 nd_low = PFN_PHYS(MAX_DMA_PFN);
196 const u64 nd_high = PFN_PHYS(max_pfn_mapped);
197 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
198 bool remapped = false;
199 u64 nd_pa;
200 void *nd;
201 int tnid;
202
203 /*
204 * Don't confuse VM with a node that doesn't have the
205 * minimum amount of memory:
206 */
207 if (end && (end - start) < NODE_MIN_SIZE)
208 return;
209
210 /* initialize remap allocator before aligning to ZONE_ALIGN */
211 init_alloc_remap(nid, start, end);
212
213 start = roundup(start, ZONE_ALIGN);
214
215 printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n",
216 nid, start, end);
217
218 /*
219 * Allocate node data. Try remap allocator first, node-local
220 * memory and then any node. Never allocate in DMA zone.
221 */
222 nd = alloc_remap(nid, nd_size);
223 if (nd) {
224 nd_pa = __pa(nd);
225 remapped = true;
226 } else {
227 nd_pa = memblock_x86_find_in_range_node(nid, nd_low, nd_high,
228 nd_size, SMP_CACHE_BYTES);
229 if (nd_pa == MEMBLOCK_ERROR)
230 nd_pa = memblock_find_in_range(nd_low, nd_high,
231 nd_size, SMP_CACHE_BYTES);
232 if (nd_pa == MEMBLOCK_ERROR) {
233 pr_err("Cannot find %zu bytes in node %d\n",
234 nd_size, nid);
235 return;
236 }
237 memblock_x86_reserve_range(nd_pa, nd_pa + nd_size, "NODE_DATA");
238 nd = __va(nd_pa);
239 }
240
241 /* report and initialize */
242 printk(KERN_INFO " NODE_DATA [%016Lx - %016Lx]%s\n",
243 nd_pa, nd_pa + nd_size - 1, remapped ? " (remapped)" : "");
244 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
245 if (!remapped && tnid != nid)
246 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
247
248 node_data[nid] = nd;
249 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
250 NODE_DATA(nid)->node_id = nid;
251 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
252 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
253
254 node_set_online(nid);
255}
256
257/**
258 * numa_cleanup_meminfo - Cleanup a numa_meminfo
259 * @mi: numa_meminfo to clean up
260 *
261 * Sanitize @mi by merging and removing unncessary memblks. Also check for
262 * conflicts and clear unused memblks.
263 *
264 * RETURNS:
265 * 0 on success, -errno on failure.
266 */
267int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
268{
269 const u64 low = 0;
270 const u64 high = PFN_PHYS(max_pfn);
271 int i, j, k;
272
273 /* first, trim all entries */
274 for (i = 0; i < mi->nr_blks; i++) {
275 struct numa_memblk *bi = &mi->blk[i];
276
277 /* make sure all blocks are inside the limits */
278 bi->start = max(bi->start, low);
279 bi->end = min(bi->end, high);
280
281 /* and there's no empty block */
282 if (bi->start >= bi->end)
283 numa_remove_memblk_from(i--, mi);
284 }
285
286 /* merge neighboring / overlapping entries */
287 for (i = 0; i < mi->nr_blks; i++) {
288 struct numa_memblk *bi = &mi->blk[i];
289
290 for (j = i + 1; j < mi->nr_blks; j++) {
291 struct numa_memblk *bj = &mi->blk[j];
292 u64 start, end;
293
294 /*
295 * See whether there are overlapping blocks. Whine
296 * about but allow overlaps of the same nid. They
297 * will be merged below.
298 */
299 if (bi->end > bj->start && bi->start < bj->end) {
300 if (bi->nid != bj->nid) {
301 pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
302 bi->nid, bi->start, bi->end,
303 bj->nid, bj->start, bj->end);
304 return -EINVAL;
305 }
306 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
307 bi->nid, bi->start, bi->end,
308 bj->start, bj->end);
309 }
310
311 /*
312 * Join together blocks on the same node, holes
313 * between which don't overlap with memory on other
314 * nodes.
315 */
316 if (bi->nid != bj->nid)
317 continue;
318 start = min(bi->start, bj->start);
319 end = max(bi->end, bj->end);
320 for (k = 0; k < mi->nr_blks; k++) {
321 struct numa_memblk *bk = &mi->blk[k];
322
323 if (bi->nid == bk->nid)
324 continue;
325 if (start < bk->end && end > bk->start)
326 break;
327 }
328 if (k < mi->nr_blks)
329 continue;
330 printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n",
331 bi->nid, bi->start, bi->end, bj->start, bj->end,
332 start, end);
333 bi->start = start;
334 bi->end = end;
335 numa_remove_memblk_from(j--, mi);
336 }
337 }
338
339 /* clear unused ones */
340 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
341 mi->blk[i].start = mi->blk[i].end = 0;
342 mi->blk[i].nid = NUMA_NO_NODE;
343 }
344
345 return 0;
346}
347
348/*
349 * Set nodes, which have memory in @mi, in *@nodemask.
350 */
351static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
352 const struct numa_meminfo *mi)
353{
354 int i;
355
356 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
357 if (mi->blk[i].start != mi->blk[i].end &&
358 mi->blk[i].nid != NUMA_NO_NODE)
359 node_set(mi->blk[i].nid, *nodemask);
360}
361
362/**
363 * numa_reset_distance - Reset NUMA distance table
364 *
365 * The current table is freed. The next numa_set_distance() call will
366 * create a new one.
367 */
368void __init numa_reset_distance(void)
369{
370 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
371
372 /* numa_distance could be 1LU marking allocation failure, test cnt */
373 if (numa_distance_cnt)
374 memblock_x86_free_range(__pa(numa_distance),
375 __pa(numa_distance) + size);
376 numa_distance_cnt = 0;
377 numa_distance = NULL; /* enable table creation */
378}
379
380static int __init numa_alloc_distance(void)
381{
382 nodemask_t nodes_parsed;
383 size_t size;
384 int i, j, cnt = 0;
385 u64 phys;
386
387 /* size the new table and allocate it */
388 nodes_parsed = numa_nodes_parsed;
389 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
390
391 for_each_node_mask(i, nodes_parsed)
392 cnt = i;
393 cnt++;
394 size = cnt * cnt * sizeof(numa_distance[0]);
395
396 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
397 size, PAGE_SIZE);
398 if (phys == MEMBLOCK_ERROR) {
399 pr_warning("NUMA: Warning: can't allocate distance table!\n");
400 /* don't retry until explicitly reset */
401 numa_distance = (void *)1LU;
402 return -ENOMEM;
403 }
404 memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
405
406 numa_distance = __va(phys);
407 numa_distance_cnt = cnt;
408
409 /* fill with the default distances */
410 for (i = 0; i < cnt; i++)
411 for (j = 0; j < cnt; j++)
412 numa_distance[i * cnt + j] = i == j ?
413 LOCAL_DISTANCE : REMOTE_DISTANCE;
414 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
415
416 return 0;
417}
418
419/**
420 * numa_set_distance - Set NUMA distance from one NUMA to another
421 * @from: the 'from' node to set distance
422 * @to: the 'to' node to set distance
423 * @distance: NUMA distance
424 *
425 * Set the distance from node @from to @to to @distance. If distance table
426 * doesn't exist, one which is large enough to accommodate all the currently
427 * known nodes will be created.
428 *
429 * If such table cannot be allocated, a warning is printed and further
430 * calls are ignored until the distance table is reset with
431 * numa_reset_distance().
432 *
433 * If @from or @to is higher than the highest known node at the time of
434 * table creation or @distance doesn't make sense, the call is ignored.
435 * This is to allow simplification of specific NUMA config implementations.
436 */
437void __init numa_set_distance(int from, int to, int distance)
438{
439 if (!numa_distance && numa_alloc_distance() < 0)
440 return;
441
442 if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
443 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
444 from, to, distance);
445 return;
446 }
447
448 if ((u8)distance != distance ||
449 (from == to && distance != LOCAL_DISTANCE)) {
450 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
451 from, to, distance);
452 return;
453 }
454
455 numa_distance[from * numa_distance_cnt + to] = distance;
456}
457
458int __node_distance(int from, int to)
459{
460 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
461 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
462 return numa_distance[from * numa_distance_cnt + to];
463}
464EXPORT_SYMBOL(__node_distance);
465
466/*
467 * Sanity check to catch more bad NUMA configurations (they are amazingly
468 * common). Make sure the nodes cover all memory.
469 */
470static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
471{
472 u64 numaram, e820ram;
473 int i;
474
475 numaram = 0;
476 for (i = 0; i < mi->nr_blks; i++) {
477 u64 s = mi->blk[i].start >> PAGE_SHIFT;
478 u64 e = mi->blk[i].end >> PAGE_SHIFT;
479 numaram += e - s;
480 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
481 if ((s64)numaram < 0)
482 numaram = 0;
483 }
484
485 e820ram = max_pfn - (memblock_x86_hole_size(0,
486 PFN_PHYS(max_pfn)) >> PAGE_SHIFT);
487 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
488 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
489 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
490 (numaram << PAGE_SHIFT) >> 20,
491 (e820ram << PAGE_SHIFT) >> 20);
492 return false;
493 }
494 return true;
495}
496
497static int __init numa_register_memblks(struct numa_meminfo *mi)
498{
499 unsigned long uninitialized_var(pfn_align);
500 int i, nid;
501
502 /* Account for nodes with cpus and no memory */
503 node_possible_map = numa_nodes_parsed;
504 numa_nodemask_from_meminfo(&node_possible_map, mi);
505 if (WARN_ON(nodes_empty(node_possible_map)))
506 return -EINVAL;
507
508 for (i = 0; i < mi->nr_blks; i++)
509 memblock_x86_register_active_regions(mi->blk[i].nid,
510 mi->blk[i].start >> PAGE_SHIFT,
511 mi->blk[i].end >> PAGE_SHIFT);
512
513 /* for out of order entries */
514 sort_node_map();
515
516 /*
517 * If sections array is gonna be used for pfn -> nid mapping, check
518 * whether its granularity is fine enough.
519 */
520#ifdef NODE_NOT_IN_PAGE_FLAGS
521 pfn_align = node_map_pfn_alignment();
522 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
523 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
524 PFN_PHYS(pfn_align) >> 20,
525 PFN_PHYS(PAGES_PER_SECTION) >> 20);
526 return -EINVAL;
527 }
528#endif
529 if (!numa_meminfo_cover_memory(mi))
530 return -EINVAL;
531
532 /* Finally register nodes. */
533 for_each_node_mask(nid, node_possible_map) {
534 u64 start = PFN_PHYS(max_pfn);
535 u64 end = 0;
536
537 for (i = 0; i < mi->nr_blks; i++) {
538 if (nid != mi->blk[i].nid)
539 continue;
540 start = min(mi->blk[i].start, start);
541 end = max(mi->blk[i].end, end);
542 }
543
544 if (start < end)
545 setup_node_data(nid, start, end);
546 }
547
548 return 0;
549}
550
551/*
552 * There are unfortunately some poorly designed mainboards around that
553 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
554 * mapping. To avoid this fill in the mapping for all possible CPUs,
555 * as the number of CPUs is not known yet. We round robin the existing
556 * nodes.
557 */
558static void __init numa_init_array(void)
559{
560 int rr, i;
561
562 rr = first_node(node_online_map);
563 for (i = 0; i < nr_cpu_ids; i++) {
564 if (early_cpu_to_node(i) != NUMA_NO_NODE)
565 continue;
566 numa_set_node(i, rr);
567 rr = next_node(rr, node_online_map);
568 if (rr == MAX_NUMNODES)
569 rr = first_node(node_online_map);
570 }
571}
572
573static int __init numa_init(int (*init_func)(void))
574{
575 int i;
576 int ret;
577
578 for (i = 0; i < MAX_LOCAL_APIC; i++)
579 set_apicid_to_node(i, NUMA_NO_NODE);
580
581 nodes_clear(numa_nodes_parsed);
582 nodes_clear(node_possible_map);
583 nodes_clear(node_online_map);
584 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
585 remove_all_active_ranges();
586 numa_reset_distance();
587
588 ret = init_func();
589 if (ret < 0)
590 return ret;
591 ret = numa_cleanup_meminfo(&numa_meminfo);
592 if (ret < 0)
593 return ret;
594
595 numa_emulation(&numa_meminfo, numa_distance_cnt);
596
597 ret = numa_register_memblks(&numa_meminfo);
598 if (ret < 0)
599 return ret;
600
601 for (i = 0; i < nr_cpu_ids; i++) {
602 int nid = early_cpu_to_node(i);
603
604 if (nid == NUMA_NO_NODE)
605 continue;
606 if (!node_online(nid))
607 numa_clear_node(i);
608 }
609 numa_init_array();
610 return 0;
611}
612
613/**
614 * dummy_numa_init - Fallback dummy NUMA init
615 *
616 * Used if there's no underlying NUMA architecture, NUMA initialization
617 * fails, or NUMA is disabled on the command line.
618 *
619 * Must online at least one node and add memory blocks that cover all
620 * allowed memory. This function must not fail.
621 */
622static int __init dummy_numa_init(void)
623{
624 printk(KERN_INFO "%s\n",
625 numa_off ? "NUMA turned off" : "No NUMA configuration found");
626 printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n",
627 0LLU, PFN_PHYS(max_pfn));
628
629 node_set(0, numa_nodes_parsed);
630 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
631
632 return 0;
633}
634
635/**
636 * x86_numa_init - Initialize NUMA
637 *
638 * Try each configured NUMA initialization method until one succeeds. The
639 * last fallback is dummy single node config encomapssing whole memory and
640 * never fails.
641 */
642void __init x86_numa_init(void)
643{
644 if (!numa_off) {
645#ifdef CONFIG_X86_NUMAQ
646 if (!numa_init(numaq_numa_init))
647 return;
648#endif
649#ifdef CONFIG_ACPI_NUMA
650 if (!numa_init(x86_acpi_numa_init))
651 return;
652#endif
653#ifdef CONFIG_AMD_NUMA
654 if (!numa_init(amd_numa_init))
655 return;
656#endif
657 }
658
659 numa_init(dummy_numa_init);
660}
661
662static __init int find_near_online_node(int node)
663{
664 int n, val;
665 int min_val = INT_MAX;
666 int best_node = -1;
667
668 for_each_online_node(n) {
669 val = node_distance(node, n);
670
671 if (val < min_val) {
672 min_val = val;
673 best_node = n;
674 }
675 }
676
677 return best_node;
678}
679
680/*
681 * Setup early cpu_to_node.
682 *
683 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
684 * and apicid_to_node[] tables have valid entries for a CPU.
685 * This means we skip cpu_to_node[] initialisation for NUMA
686 * emulation and faking node case (when running a kernel compiled
687 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
688 * is already initialized in a round robin manner at numa_init_array,
689 * prior to this call, and this initialization is good enough
690 * for the fake NUMA cases.
691 *
692 * Called before the per_cpu areas are setup.
693 */
694void __init init_cpu_to_node(void)
695{
696 int cpu;
697 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
698
699 BUG_ON(cpu_to_apicid == NULL);
700
701 for_each_possible_cpu(cpu) {
702 int node = numa_cpu_node(cpu);
703
704 if (node == NUMA_NO_NODE)
705 continue;
706 if (!node_online(node))
707 node = find_near_online_node(node);
708 numa_set_node(cpu, node);
709 }
710}
711
712#ifndef CONFIG_DEBUG_PER_CPU_MAPS
713
714# ifndef CONFIG_NUMA_EMU
715void __cpuinit numa_add_cpu(int cpu)
716{
717 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
718}
719
720void __cpuinit numa_remove_cpu(int cpu)
721{
722 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
723}
724# endif /* !CONFIG_NUMA_EMU */
725
726#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
727
728int __cpu_to_node(int cpu)
729{
730 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
731 printk(KERN_WARNING
732 "cpu_to_node(%d): usage too early!\n", cpu);
733 dump_stack();
734 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
735 }
736 return per_cpu(x86_cpu_to_node_map, cpu);
737}
738EXPORT_SYMBOL(__cpu_to_node);
739
740/*
741 * Same function as cpu_to_node() but used if called before the
742 * per_cpu areas are setup.
743 */
744int early_cpu_to_node(int cpu)
745{
746 if (early_per_cpu_ptr(x86_cpu_to_node_map))
747 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
748
749 if (!cpu_possible(cpu)) {
750 printk(KERN_WARNING
751 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
752 dump_stack();
753 return NUMA_NO_NODE;
754 }
755 return per_cpu(x86_cpu_to_node_map, cpu);
756}
757
758void debug_cpumask_set_cpu(int cpu, int node, bool enable)
759{
760 struct cpumask *mask;
761 char buf[64];
762
763 if (node == NUMA_NO_NODE) {
764 /* early_cpu_to_node() already emits a warning and trace */
765 return;
766 }
767 mask = node_to_cpumask_map[node];
768 if (!mask) {
769 pr_err("node_to_cpumask_map[%i] NULL\n", node);
770 dump_stack();
771 return;
772 }
773
774 if (enable)
775 cpumask_set_cpu(cpu, mask);
776 else
777 cpumask_clear_cpu(cpu, mask);
778
779 cpulist_scnprintf(buf, sizeof(buf), mask);
780 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
781 enable ? "numa_add_cpu" : "numa_remove_cpu",
782 cpu, node, buf);
783 return;
784}
785
786# ifndef CONFIG_NUMA_EMU
787static void __cpuinit numa_set_cpumask(int cpu, bool enable)
788{
789 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
790}
791
792void __cpuinit numa_add_cpu(int cpu)
793{
794 numa_set_cpumask(cpu, true);
795}
796
797void __cpuinit numa_remove_cpu(int cpu)
798{
799 numa_set_cpumask(cpu, false);
800}
801# endif /* !CONFIG_NUMA_EMU */
802
803/*
804 * Returns a pointer to the bitmask of CPUs on Node 'node'.
805 */
806const struct cpumask *cpumask_of_node(int node)
807{
808 if (node >= nr_node_ids) {
809 printk(KERN_WARNING
810 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
811 node, nr_node_ids);
812 dump_stack();
813 return cpu_none_mask;
814 }
815 if (node_to_cpumask_map[node] == NULL) {
816 printk(KERN_WARNING
817 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
818 node);
819 dump_stack();
820 return cpu_online_mask;
821 }
822 return node_to_cpumask_map[node];
823}
824EXPORT_SYMBOL(cpumask_of_node);
825
826#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
827
828#ifdef CONFIG_MEMORY_HOTPLUG
829int memory_add_physaddr_to_nid(u64 start)
830{
831 struct numa_meminfo *mi = &numa_meminfo;
832 int nid = mi->blk[0].nid;
833 int i;
834
835 for (i = 0; i < mi->nr_blks; i++)
836 if (mi->blk[i].start <= start && mi->blk[i].end > start)
837 nid = mi->blk[i].nid;
838 return nid;
839}
840EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
841#endif
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#ifdef CONFIG_NUMA_EMU
41 if (!strncmp(opt, "fake=", 5))
42 numa_emu_cmdline(opt + 5);
43#endif
44#ifdef CONFIG_ACPI_NUMA
45 if (!strncmp(opt, "noacpi", 6))
46 acpi_numa = -1;
47#endif
48 return 0;
49}
50early_param("numa", numa_setup);
51
52/*
53 * apicid, cpu, node mappings
54 */
55s16 __apicid_to_node[MAX_LOCAL_APIC] = {
56 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
57};
58
59int numa_cpu_node(int cpu)
60{
61 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
62
63 if (apicid != BAD_APICID)
64 return __apicid_to_node[apicid];
65 return NUMA_NO_NODE;
66}
67
68cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
69EXPORT_SYMBOL(node_to_cpumask_map);
70
71/*
72 * Map cpu index to node index
73 */
74DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
75EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
76
77void numa_set_node(int cpu, int node)
78{
79 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
80
81 /* early setting, no percpu area yet */
82 if (cpu_to_node_map) {
83 cpu_to_node_map[cpu] = node;
84 return;
85 }
86
87#ifdef CONFIG_DEBUG_PER_CPU_MAPS
88 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
89 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
90 dump_stack();
91 return;
92 }
93#endif
94 per_cpu(x86_cpu_to_node_map, cpu) = node;
95
96 set_cpu_numa_node(cpu, node);
97}
98
99void numa_clear_node(int cpu)
100{
101 numa_set_node(cpu, NUMA_NO_NODE);
102}
103
104/*
105 * Allocate node_to_cpumask_map based on number of available nodes
106 * Requires node_possible_map to be valid.
107 *
108 * Note: cpumask_of_node() is not valid until after this is done.
109 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
110 */
111void __init setup_node_to_cpumask_map(void)
112{
113 unsigned int node;
114
115 /* setup nr_node_ids if not done yet */
116 if (nr_node_ids == MAX_NUMNODES)
117 setup_nr_node_ids();
118
119 /* allocate the map */
120 for (node = 0; node < nr_node_ids; node++)
121 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
122
123 /* cpumask_of_node() will now work */
124 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
125}
126
127static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
128 struct numa_meminfo *mi)
129{
130 /* ignore zero length blks */
131 if (start == end)
132 return 0;
133
134 /* whine about and ignore invalid blks */
135 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
136 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
137 nid, start, end - 1);
138 return 0;
139 }
140
141 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
142 pr_err("too many memblk ranges\n");
143 return -EINVAL;
144 }
145
146 mi->blk[mi->nr_blks].start = start;
147 mi->blk[mi->nr_blks].end = end;
148 mi->blk[mi->nr_blks].nid = nid;
149 mi->nr_blks++;
150 return 0;
151}
152
153/**
154 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
155 * @idx: Index of memblk to remove
156 * @mi: numa_meminfo to remove memblk from
157 *
158 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
159 * decrementing @mi->nr_blks.
160 */
161void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
162{
163 mi->nr_blks--;
164 memmove(&mi->blk[idx], &mi->blk[idx + 1],
165 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
166}
167
168/**
169 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another
170 * @dst: numa_meminfo to append block to
171 * @idx: Index of memblk to remove
172 * @src: numa_meminfo to remove memblk from
173 */
174static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
175 struct numa_meminfo *src)
176{
177 dst->blk[dst->nr_blks++] = src->blk[idx];
178 numa_remove_memblk_from(idx, src);
179}
180
181/**
182 * numa_add_memblk - Add one numa_memblk to numa_meminfo
183 * @nid: NUMA node ID of the new memblk
184 * @start: Start address of the new memblk
185 * @end: End address of the new memblk
186 *
187 * Add a new memblk to the default numa_meminfo.
188 *
189 * RETURNS:
190 * 0 on success, -errno on failure.
191 */
192int __init numa_add_memblk(int nid, u64 start, u64 end)
193{
194 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
195}
196
197/* Allocate NODE_DATA for a node on the local memory */
198static void __init alloc_node_data(int nid)
199{
200 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
201 u64 nd_pa;
202 void *nd;
203 int tnid;
204
205 /*
206 * Allocate node data. Try node-local memory and then any node.
207 * Never allocate in DMA zone.
208 */
209 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
210 if (!nd_pa) {
211 pr_err("Cannot find %zu bytes in any node (initial node: %d)\n",
212 nd_size, nid);
213 return;
214 }
215 nd = __va(nd_pa);
216
217 /* report and initialize */
218 printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
219 nd_pa, nd_pa + nd_size - 1);
220 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
221 if (tnid != nid)
222 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
223
224 node_data[nid] = nd;
225 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
226
227 node_set_online(nid);
228}
229
230/**
231 * numa_cleanup_meminfo - Cleanup a numa_meminfo
232 * @mi: numa_meminfo to clean up
233 *
234 * Sanitize @mi by merging and removing unnecessary memblks. Also check for
235 * conflicts and clear unused memblks.
236 *
237 * RETURNS:
238 * 0 on success, -errno on failure.
239 */
240int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
241{
242 const u64 low = 0;
243 const u64 high = PFN_PHYS(max_pfn);
244 int i, j, k;
245
246 /* first, trim all entries */
247 for (i = 0; i < mi->nr_blks; i++) {
248 struct numa_memblk *bi = &mi->blk[i];
249
250 /* move / save reserved memory ranges */
251 if (!memblock_overlaps_region(&memblock.memory,
252 bi->start, bi->end - bi->start)) {
253 numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi);
254 continue;
255 }
256
257 /* make sure all non-reserved blocks are inside the limits */
258 bi->start = max(bi->start, low);
259 bi->end = min(bi->end, high);
260
261 /* and there's no empty block */
262 if (bi->start >= bi->end)
263 numa_remove_memblk_from(i--, mi);
264 }
265
266 /* merge neighboring / overlapping entries */
267 for (i = 0; i < mi->nr_blks; i++) {
268 struct numa_memblk *bi = &mi->blk[i];
269
270 for (j = i + 1; j < mi->nr_blks; j++) {
271 struct numa_memblk *bj = &mi->blk[j];
272 u64 start, end;
273
274 /*
275 * See whether there are overlapping blocks. Whine
276 * about but allow overlaps of the same nid. They
277 * will be merged below.
278 */
279 if (bi->end > bj->start && bi->start < bj->end) {
280 if (bi->nid != bj->nid) {
281 pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
282 bi->nid, bi->start, bi->end - 1,
283 bj->nid, bj->start, bj->end - 1);
284 return -EINVAL;
285 }
286 pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
287 bi->nid, bi->start, bi->end - 1,
288 bj->start, bj->end - 1);
289 }
290
291 /*
292 * Join together blocks on the same node, holes
293 * between which don't overlap with memory on other
294 * nodes.
295 */
296 if (bi->nid != bj->nid)
297 continue;
298 start = min(bi->start, bj->start);
299 end = max(bi->end, bj->end);
300 for (k = 0; k < mi->nr_blks; k++) {
301 struct numa_memblk *bk = &mi->blk[k];
302
303 if (bi->nid == bk->nid)
304 continue;
305 if (start < bk->end && end > bk->start)
306 break;
307 }
308 if (k < mi->nr_blks)
309 continue;
310 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
311 bi->nid, bi->start, bi->end - 1, bj->start,
312 bj->end - 1, start, end - 1);
313 bi->start = start;
314 bi->end = end;
315 numa_remove_memblk_from(j--, mi);
316 }
317 }
318
319 /* clear unused ones */
320 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
321 mi->blk[i].start = mi->blk[i].end = 0;
322 mi->blk[i].nid = NUMA_NO_NODE;
323 }
324
325 return 0;
326}
327
328/*
329 * Set nodes, which have memory in @mi, in *@nodemask.
330 */
331static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
332 const struct numa_meminfo *mi)
333{
334 int i;
335
336 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
337 if (mi->blk[i].start != mi->blk[i].end &&
338 mi->blk[i].nid != NUMA_NO_NODE)
339 node_set(mi->blk[i].nid, *nodemask);
340}
341
342/**
343 * numa_reset_distance - Reset NUMA distance table
344 *
345 * The current table is freed. The next numa_set_distance() call will
346 * create a new one.
347 */
348void __init numa_reset_distance(void)
349{
350 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
351
352 /* numa_distance could be 1LU marking allocation failure, test cnt */
353 if (numa_distance_cnt)
354 memblock_free(__pa(numa_distance), size);
355 numa_distance_cnt = 0;
356 numa_distance = NULL; /* enable table creation */
357}
358
359static int __init numa_alloc_distance(void)
360{
361 nodemask_t nodes_parsed;
362 size_t size;
363 int i, j, cnt = 0;
364 u64 phys;
365
366 /* size the new table and allocate it */
367 nodes_parsed = numa_nodes_parsed;
368 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
369
370 for_each_node_mask(i, nodes_parsed)
371 cnt = i;
372 cnt++;
373 size = cnt * cnt * sizeof(numa_distance[0]);
374
375 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
376 size, PAGE_SIZE);
377 if (!phys) {
378 pr_warn("Warning: can't allocate distance table!\n");
379 /* don't retry until explicitly reset */
380 numa_distance = (void *)1LU;
381 return -ENOMEM;
382 }
383 memblock_reserve(phys, size);
384
385 numa_distance = __va(phys);
386 numa_distance_cnt = cnt;
387
388 /* fill with the default distances */
389 for (i = 0; i < cnt; i++)
390 for (j = 0; j < cnt; j++)
391 numa_distance[i * cnt + j] = i == j ?
392 LOCAL_DISTANCE : REMOTE_DISTANCE;
393 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
394
395 return 0;
396}
397
398/**
399 * numa_set_distance - Set NUMA distance from one NUMA to another
400 * @from: the 'from' node to set distance
401 * @to: the 'to' node to set distance
402 * @distance: NUMA distance
403 *
404 * Set the distance from node @from to @to to @distance. If distance table
405 * doesn't exist, one which is large enough to accommodate all the currently
406 * known nodes will be created.
407 *
408 * If such table cannot be allocated, a warning is printed and further
409 * calls are ignored until the distance table is reset with
410 * numa_reset_distance().
411 *
412 * If @from or @to is higher than the highest known node or lower than zero
413 * at the time of table creation or @distance doesn't make sense, the call
414 * is ignored.
415 * This is to allow simplification of specific NUMA config implementations.
416 */
417void __init numa_set_distance(int from, int to, int distance)
418{
419 if (!numa_distance && numa_alloc_distance() < 0)
420 return;
421
422 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
423 from < 0 || to < 0) {
424 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
425 from, to, distance);
426 return;
427 }
428
429 if ((u8)distance != distance ||
430 (from == to && distance != LOCAL_DISTANCE)) {
431 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
432 from, to, distance);
433 return;
434 }
435
436 numa_distance[from * numa_distance_cnt + to] = distance;
437}
438
439int __node_distance(int from, int to)
440{
441 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
442 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
443 return numa_distance[from * numa_distance_cnt + to];
444}
445EXPORT_SYMBOL(__node_distance);
446
447/*
448 * Sanity check to catch more bad NUMA configurations (they are amazingly
449 * common). Make sure the nodes cover all memory.
450 */
451static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
452{
453 u64 numaram, e820ram;
454 int i;
455
456 numaram = 0;
457 for (i = 0; i < mi->nr_blks; i++) {
458 u64 s = mi->blk[i].start >> PAGE_SHIFT;
459 u64 e = mi->blk[i].end >> PAGE_SHIFT;
460 numaram += e - s;
461 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
462 if ((s64)numaram < 0)
463 numaram = 0;
464 }
465
466 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
467
468 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
469 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
470 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
471 (numaram << PAGE_SHIFT) >> 20,
472 (e820ram << PAGE_SHIFT) >> 20);
473 return false;
474 }
475 return true;
476}
477
478/*
479 * Mark all currently memblock-reserved physical memory (which covers the
480 * kernel's own memory ranges) as hot-unswappable.
481 */
482static void __init numa_clear_kernel_node_hotplug(void)
483{
484 nodemask_t reserved_nodemask = NODE_MASK_NONE;
485 struct memblock_region *mb_region;
486 int i;
487
488 /*
489 * We have to do some preprocessing of memblock regions, to
490 * make them suitable for reservation.
491 *
492 * At this time, all memory regions reserved by memblock are
493 * used by the kernel, but those regions are not split up
494 * along node boundaries yet, and don't necessarily have their
495 * node ID set yet either.
496 *
497 * So iterate over all memory known to the x86 architecture,
498 * and use those ranges to set the nid in memblock.reserved.
499 * This will split up the memblock regions along node
500 * boundaries and will set the node IDs as well.
501 */
502 for (i = 0; i < numa_meminfo.nr_blks; i++) {
503 struct numa_memblk *mb = numa_meminfo.blk + i;
504 int ret;
505
506 ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
507 WARN_ON_ONCE(ret);
508 }
509
510 /*
511 * Now go over all reserved memblock regions, to construct a
512 * node mask of all kernel reserved memory areas.
513 *
514 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
515 * numa_meminfo might not include all memblock.reserved
516 * memory ranges, because quirks such as trim_snb_memory()
517 * reserve specific pages for Sandy Bridge graphics. ]
518 */
519 for_each_memblock(reserved, mb_region) {
520 int nid = memblock_get_region_node(mb_region);
521
522 if (nid != MAX_NUMNODES)
523 node_set(nid, reserved_nodemask);
524 }
525
526 /*
527 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
528 * belonging to the reserved node mask.
529 *
530 * Note that this will include memory regions that reside
531 * on nodes that contain kernel memory - entire nodes
532 * become hot-unpluggable:
533 */
534 for (i = 0; i < numa_meminfo.nr_blks; i++) {
535 struct numa_memblk *mb = numa_meminfo.blk + i;
536
537 if (!node_isset(mb->nid, reserved_nodemask))
538 continue;
539
540 memblock_clear_hotplug(mb->start, mb->end - mb->start);
541 }
542}
543
544static int __init numa_register_memblks(struct numa_meminfo *mi)
545{
546 int i, nid;
547
548 /* Account for nodes with cpus and no memory */
549 node_possible_map = numa_nodes_parsed;
550 numa_nodemask_from_meminfo(&node_possible_map, mi);
551 if (WARN_ON(nodes_empty(node_possible_map)))
552 return -EINVAL;
553
554 for (i = 0; i < mi->nr_blks; i++) {
555 struct numa_memblk *mb = &mi->blk[i];
556 memblock_set_node(mb->start, mb->end - mb->start,
557 &memblock.memory, mb->nid);
558 }
559
560 /*
561 * At very early time, the kernel have to use some memory such as
562 * loading the kernel image. We cannot prevent this anyway. So any
563 * node the kernel resides in should be un-hotpluggable.
564 *
565 * And when we come here, alloc node data won't fail.
566 */
567 numa_clear_kernel_node_hotplug();
568
569 /*
570 * If sections array is gonna be used for pfn -> nid mapping, check
571 * whether its granularity is fine enough.
572 */
573 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
574 unsigned long pfn_align = node_map_pfn_alignment();
575
576 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
577 pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
578 PFN_PHYS(pfn_align) >> 20,
579 PFN_PHYS(PAGES_PER_SECTION) >> 20);
580 return -EINVAL;
581 }
582 }
583 if (!numa_meminfo_cover_memory(mi))
584 return -EINVAL;
585
586 /* Finally register nodes. */
587 for_each_node_mask(nid, node_possible_map) {
588 u64 start = PFN_PHYS(max_pfn);
589 u64 end = 0;
590
591 for (i = 0; i < mi->nr_blks; i++) {
592 if (nid != mi->blk[i].nid)
593 continue;
594 start = min(mi->blk[i].start, start);
595 end = max(mi->blk[i].end, end);
596 }
597
598 if (start >= end)
599 continue;
600
601 /*
602 * Don't confuse VM with a node that doesn't have the
603 * minimum amount of memory:
604 */
605 if (end && (end - start) < NODE_MIN_SIZE)
606 continue;
607
608 alloc_node_data(nid);
609 }
610
611 /* Dump memblock with node info and return. */
612 memblock_dump_all();
613 return 0;
614}
615
616/*
617 * There are unfortunately some poorly designed mainboards around that
618 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
619 * mapping. To avoid this fill in the mapping for all possible CPUs,
620 * as the number of CPUs is not known yet. We round robin the existing
621 * nodes.
622 */
623static void __init numa_init_array(void)
624{
625 int rr, i;
626
627 rr = first_node(node_online_map);
628 for (i = 0; i < nr_cpu_ids; i++) {
629 if (early_cpu_to_node(i) != NUMA_NO_NODE)
630 continue;
631 numa_set_node(i, rr);
632 rr = next_node_in(rr, node_online_map);
633 }
634}
635
636static int __init numa_init(int (*init_func)(void))
637{
638 int i;
639 int ret;
640
641 for (i = 0; i < MAX_LOCAL_APIC; i++)
642 set_apicid_to_node(i, NUMA_NO_NODE);
643
644 nodes_clear(numa_nodes_parsed);
645 nodes_clear(node_possible_map);
646 nodes_clear(node_online_map);
647 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
648 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
649 MAX_NUMNODES));
650 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
651 MAX_NUMNODES));
652 /* In case that parsing SRAT failed. */
653 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
654 numa_reset_distance();
655
656 ret = init_func();
657 if (ret < 0)
658 return ret;
659
660 /*
661 * We reset memblock back to the top-down direction
662 * here because if we configured ACPI_NUMA, we have
663 * parsed SRAT in init_func(). It is ok to have the
664 * reset here even if we did't configure ACPI_NUMA
665 * or acpi numa init fails and fallbacks to dummy
666 * numa init.
667 */
668 memblock_set_bottom_up(false);
669
670 ret = numa_cleanup_meminfo(&numa_meminfo);
671 if (ret < 0)
672 return ret;
673
674 numa_emulation(&numa_meminfo, numa_distance_cnt);
675
676 ret = numa_register_memblks(&numa_meminfo);
677 if (ret < 0)
678 return ret;
679
680 for (i = 0; i < nr_cpu_ids; i++) {
681 int nid = early_cpu_to_node(i);
682
683 if (nid == NUMA_NO_NODE)
684 continue;
685 if (!node_online(nid))
686 numa_clear_node(i);
687 }
688 numa_init_array();
689
690 return 0;
691}
692
693/**
694 * dummy_numa_init - Fallback dummy NUMA init
695 *
696 * Used if there's no underlying NUMA architecture, NUMA initialization
697 * fails, or NUMA is disabled on the command line.
698 *
699 * Must online at least one node and add memory blocks that cover all
700 * allowed memory. This function must not fail.
701 */
702static int __init dummy_numa_init(void)
703{
704 printk(KERN_INFO "%s\n",
705 numa_off ? "NUMA turned off" : "No NUMA configuration found");
706 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
707 0LLU, PFN_PHYS(max_pfn) - 1);
708
709 node_set(0, numa_nodes_parsed);
710 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
711
712 return 0;
713}
714
715/**
716 * x86_numa_init - Initialize NUMA
717 *
718 * Try each configured NUMA initialization method until one succeeds. The
719 * last fallback is dummy single node config encompassing whole memory and
720 * never fails.
721 */
722void __init x86_numa_init(void)
723{
724 if (!numa_off) {
725#ifdef CONFIG_ACPI_NUMA
726 if (!numa_init(x86_acpi_numa_init))
727 return;
728#endif
729#ifdef CONFIG_AMD_NUMA
730 if (!numa_init(amd_numa_init))
731 return;
732#endif
733 }
734
735 numa_init(dummy_numa_init);
736}
737
738static void __init init_memory_less_node(int nid)
739{
740 /* Allocate and initialize node data. Memory-less node is now online.*/
741 alloc_node_data(nid);
742 free_area_init_memoryless_node(nid);
743
744 /*
745 * All zonelists will be built later in start_kernel() after per cpu
746 * areas are initialized.
747 */
748}
749
750/*
751 * Setup early cpu_to_node.
752 *
753 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
754 * and apicid_to_node[] tables have valid entries for a CPU.
755 * This means we skip cpu_to_node[] initialisation for NUMA
756 * emulation and faking node case (when running a kernel compiled
757 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
758 * is already initialized in a round robin manner at numa_init_array,
759 * prior to this call, and this initialization is good enough
760 * for the fake NUMA cases.
761 *
762 * Called before the per_cpu areas are setup.
763 */
764void __init init_cpu_to_node(void)
765{
766 int cpu;
767 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
768
769 BUG_ON(cpu_to_apicid == NULL);
770
771 for_each_possible_cpu(cpu) {
772 int node = numa_cpu_node(cpu);
773
774 if (node == NUMA_NO_NODE)
775 continue;
776
777 if (!node_online(node))
778 init_memory_less_node(node);
779
780 numa_set_node(cpu, node);
781 }
782}
783
784#ifndef CONFIG_DEBUG_PER_CPU_MAPS
785
786# ifndef CONFIG_NUMA_EMU
787void numa_add_cpu(int cpu)
788{
789 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
790}
791
792void numa_remove_cpu(int cpu)
793{
794 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
795}
796# endif /* !CONFIG_NUMA_EMU */
797
798#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
799
800int __cpu_to_node(int cpu)
801{
802 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
803 printk(KERN_WARNING
804 "cpu_to_node(%d): usage too early!\n", cpu);
805 dump_stack();
806 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
807 }
808 return per_cpu(x86_cpu_to_node_map, cpu);
809}
810EXPORT_SYMBOL(__cpu_to_node);
811
812/*
813 * Same function as cpu_to_node() but used if called before the
814 * per_cpu areas are setup.
815 */
816int early_cpu_to_node(int cpu)
817{
818 if (early_per_cpu_ptr(x86_cpu_to_node_map))
819 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
820
821 if (!cpu_possible(cpu)) {
822 printk(KERN_WARNING
823 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
824 dump_stack();
825 return NUMA_NO_NODE;
826 }
827 return per_cpu(x86_cpu_to_node_map, cpu);
828}
829
830void debug_cpumask_set_cpu(int cpu, int node, bool enable)
831{
832 struct cpumask *mask;
833
834 if (node == NUMA_NO_NODE) {
835 /* early_cpu_to_node() already emits a warning and trace */
836 return;
837 }
838 mask = node_to_cpumask_map[node];
839 if (!mask) {
840 pr_err("node_to_cpumask_map[%i] NULL\n", node);
841 dump_stack();
842 return;
843 }
844
845 if (enable)
846 cpumask_set_cpu(cpu, mask);
847 else
848 cpumask_clear_cpu(cpu, mask);
849
850 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
851 enable ? "numa_add_cpu" : "numa_remove_cpu",
852 cpu, node, cpumask_pr_args(mask));
853 return;
854}
855
856# ifndef CONFIG_NUMA_EMU
857static void numa_set_cpumask(int cpu, bool enable)
858{
859 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
860}
861
862void numa_add_cpu(int cpu)
863{
864 numa_set_cpumask(cpu, true);
865}
866
867void numa_remove_cpu(int cpu)
868{
869 numa_set_cpumask(cpu, false);
870}
871# endif /* !CONFIG_NUMA_EMU */
872
873/*
874 * Returns a pointer to the bitmask of CPUs on Node 'node'.
875 */
876const struct cpumask *cpumask_of_node(int node)
877{
878 if ((unsigned)node >= nr_node_ids) {
879 printk(KERN_WARNING
880 "cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
881 node, nr_node_ids);
882 dump_stack();
883 return cpu_none_mask;
884 }
885 if (node_to_cpumask_map[node] == NULL) {
886 printk(KERN_WARNING
887 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
888 node);
889 dump_stack();
890 return cpu_online_mask;
891 }
892 return node_to_cpumask_map[node];
893}
894EXPORT_SYMBOL(cpumask_of_node);
895
896#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
897
898#ifdef CONFIG_NUMA_KEEP_MEMINFO
899static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
900{
901 int i;
902
903 for (i = 0; i < mi->nr_blks; i++)
904 if (mi->blk[i].start <= start && mi->blk[i].end > start)
905 return mi->blk[i].nid;
906 return NUMA_NO_NODE;
907}
908
909int phys_to_target_node(phys_addr_t start)
910{
911 int nid = meminfo_to_nid(&numa_meminfo, start);
912
913 /*
914 * Prefer online nodes, but if reserved memory might be
915 * hot-added continue the search with reserved ranges.
916 */
917 if (nid != NUMA_NO_NODE)
918 return nid;
919
920 return meminfo_to_nid(&numa_reserved_meminfo, start);
921}
922EXPORT_SYMBOL_GPL(phys_to_target_node);
923
924int memory_add_physaddr_to_nid(u64 start)
925{
926 int nid = meminfo_to_nid(&numa_meminfo, start);
927
928 if (nid == NUMA_NO_NODE)
929 nid = numa_meminfo.blk[0].nid;
930 return nid;
931}
932#endif