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