1
2
3
4
5#include <linux/kernel.h>
6#include <linux/mm.h>
7#include <linux/string.h>
8#include <linux/init.h>
9#include <linux/bootmem.h>
10#include <linux/mmzone.h>
11#include <linux/ctype.h>
12#include <linux/module.h>
13#include <linux/nodemask.h>
14#include <linux/sched.h>
15
16#include <asm/e820.h>
17#include <asm/proto.h>
18#include <asm/dma.h>
19#include <asm/numa.h>
20#include <asm/acpi.h>
21#include <asm/k8.h>
22
23struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
24EXPORT_SYMBOL(node_data);
25
26struct memnode memnode;
27
28s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
29 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
30};
31
32int numa_off __initdata;
33static unsigned long __initdata nodemap_addr;
34static unsigned long __initdata nodemap_size;
35
36DEFINE_PER_CPU(int, node_number) = 0;
37EXPORT_PER_CPU_SYMBOL(node_number);
38
39
40
41
42DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
43EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
44
45
46
47
48
49
50
51
52static int __init populate_memnodemap(const struct bootnode *nodes,
53 int numnodes, int shift, int *nodeids)
54{
55 unsigned long addr, end;
56 int i, res = -1;
57
58 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
59 for (i = 0; i < numnodes; i++) {
60 addr = nodes[i].start;
61 end = nodes[i].end;
62 if (addr >= end)
63 continue;
64 if ((end >> shift) >= memnodemapsize)
65 return 0;
66 do {
67 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
68 return -1;
69
70 if (!nodeids)
71 memnodemap[addr >> shift] = i;
72 else
73 memnodemap[addr >> shift] = nodeids[i];
74
75 addr += (1UL << shift);
76 } while (addr < end);
77 res = 1;
78 }
79 return res;
80}
81
82static int __init allocate_cachealigned_memnodemap(void)
83{
84 unsigned long addr;
85
86 memnodemap = memnode.embedded_map;
87 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
88 return 0;
89
90 addr = 0x8000;
91 nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
92 nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT,
93 nodemap_size, L1_CACHE_BYTES);
94 if (nodemap_addr == -1UL) {
95 printk(KERN_ERR
96 "NUMA: Unable to allocate Memory to Node hash map\n");
97 nodemap_addr = nodemap_size = 0;
98 return -1;
99 }
100 memnodemap = phys_to_virt(nodemap_addr);
101 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
102
103 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
104 nodemap_addr, nodemap_addr + nodemap_size);
105 return 0;
106}
107
108
109
110
111
112static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
113 int numnodes)
114{
115 int i, nodes_used = 0;
116 unsigned long start, end;
117 unsigned long bitfield = 0, memtop = 0;
118
119 for (i = 0; i < numnodes; i++) {
120 start = nodes[i].start;
121 end = nodes[i].end;
122 if (start >= end)
123 continue;
124 bitfield |= start;
125 nodes_used++;
126 if (end > memtop)
127 memtop = end;
128 }
129 if (nodes_used <= 1)
130 i = 63;
131 else
132 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
133 memnodemapsize = (memtop >> i)+1;
134 return i;
135}
136
137int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
138 int *nodeids)
139{
140 int shift;
141
142 shift = extract_lsb_from_nodes(nodes, numnodes);
143 if (allocate_cachealigned_memnodemap())
144 return -1;
145 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
146 shift);
147
148 if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
149 printk(KERN_INFO "Your memory is not aligned you need to "
150 "rebuild your kernel with a bigger NODEMAPSIZE "
151 "shift=%d\n", shift);
152 return -1;
153 }
154 return shift;
155}
156
157int __meminit __early_pfn_to_nid(unsigned long pfn)
158{
159 return phys_to_nid(pfn << PAGE_SHIFT);
160}
161
162static void * __init early_node_mem(int nodeid, unsigned long start,
163 unsigned long end, unsigned long size,
164 unsigned long align)
165{
166 unsigned long mem = find_e820_area(start, end, size, align);
167 void *ptr;
168
169 if (mem != -1L)
170 return __va(mem);
171
172 ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
173 if (ptr == NULL) {
174 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
175 size, nodeid);
176 return NULL;
177 }
178 return ptr;
179}
180
181
182void __init
183setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
184{
185 unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size;
186 const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
187 unsigned long bootmap_start, nodedata_phys;
188 void *bootmap;
189 int nid;
190
191 if (!end)
192 return;
193
194
195
196
197
198 if (end && (end - start) < NODE_MIN_SIZE)
199 return;
200
201 start = roundup(start, ZONE_ALIGN);
202
203 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
204 start, end);
205
206 start_pfn = start >> PAGE_SHIFT;
207 last_pfn = end >> PAGE_SHIFT;
208
209 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
210 SMP_CACHE_BYTES);
211 if (node_data[nodeid] == NULL)
212 return;
213 nodedata_phys = __pa(node_data[nodeid]);
214 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
215 nodedata_phys + pgdat_size - 1);
216
217 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
218 NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid];
219 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
220 NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
221
222
223
224
225
226
227
228
229 bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn);
230 nid = phys_to_nid(nodedata_phys);
231 if (nid == nodeid)
232 bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE);
233 else
234 bootmap_start = roundup(start, PAGE_SIZE);
235
236
237
238
239 bootmap = early_node_mem(nodeid, bootmap_start, end,
240 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
241 if (bootmap == NULL) {
242 if (nodedata_phys < start || nodedata_phys >= end)
243 free_bootmem(nodedata_phys, pgdat_size);
244 node_data[nodeid] = NULL;
245 return;
246 }
247 bootmap_start = __pa(bootmap);
248
249 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
250 bootmap_start >> PAGE_SHIFT,
251 start_pfn, last_pfn);
252
253 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
254 bootmap_start, bootmap_start + bootmap_size - 1,
255 bootmap_pages);
256
257 free_bootmem_with_active_regions(nodeid, end);
258
259
260
261
262
263
264 early_res_to_bootmem(start, end);
265
266
267
268
269
270 if (nid != nodeid)
271 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
272 else
273 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys,
274 pgdat_size, BOOTMEM_DEFAULT);
275 nid = phys_to_nid(bootmap_start);
276 if (nid != nodeid)
277 printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid);
278 else
279 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
280 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
281
282 node_set_online(nodeid);
283}
284
285
286
287
288
289
290
291
292void __init numa_init_array(void)
293{
294 int rr, i;
295
296 rr = first_node(node_online_map);
297 for (i = 0; i < nr_cpu_ids; i++) {
298 if (early_cpu_to_node(i) != NUMA_NO_NODE)
299 continue;
300 numa_set_node(i, rr);
301 rr = next_node(rr, node_online_map);
302 if (rr == MAX_NUMNODES)
303 rr = first_node(node_online_map);
304 }
305}
306
307#ifdef CONFIG_NUMA_EMU
308
309static char *cmdline __initdata;
310
311
312
313
314
315
316
317
318static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
319 u64 size, u64 max_addr)
320{
321 int ret = 0;
322
323 nodes[nid].start = *addr;
324 *addr += size;
325 if (*addr >= max_addr) {
326 *addr = max_addr;
327 ret = -1;
328 }
329 nodes[nid].end = *addr;
330 node_set(nid, node_possible_map);
331 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
332 nodes[nid].start, nodes[nid].end,
333 (nodes[nid].end - nodes[nid].start) >> 20);
334 return ret;
335}
336
337
338
339
340
341
342static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
343 u64 max_addr, int node_start,
344 int num_nodes)
345{
346 unsigned int big;
347 u64 size;
348 int i;
349
350 if (num_nodes <= 0)
351 return -1;
352 if (num_nodes > MAX_NUMNODES)
353 num_nodes = MAX_NUMNODES;
354 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
355 num_nodes;
356
357
358
359
360 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
361 FAKE_NODE_MIN_SIZE;
362
363
364 size &= FAKE_NODE_MIN_HASH_MASK;
365 if (!size) {
366 printk(KERN_ERR "Not enough memory for each node. "
367 "NUMA emulation disabled.\n");
368 return -1;
369 }
370
371 for (i = node_start; i < num_nodes + node_start; i++) {
372 u64 end = *addr + size;
373
374 if (i < big)
375 end += FAKE_NODE_MIN_SIZE;
376
377
378
379
380 if (i == num_nodes + node_start - 1)
381 end = max_addr;
382 else
383 while (end - *addr - e820_hole_size(*addr, end) <
384 size) {
385 end += FAKE_NODE_MIN_SIZE;
386 if (end > max_addr) {
387 end = max_addr;
388 break;
389 }
390 }
391 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
392 break;
393 }
394 return i - node_start + 1;
395}
396
397
398
399
400
401
402static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
403 u64 max_addr, int node_start, u64 size)
404{
405 int i = node_start;
406 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
407 while (!setup_node_range(i++, nodes, addr, size, max_addr))
408 ;
409 return i - node_start;
410}
411
412
413
414
415
416static struct bootnode nodes[MAX_NUMNODES] __initdata;
417
418static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn)
419{
420 u64 size, addr = start_pfn << PAGE_SHIFT;
421 u64 max_addr = last_pfn << PAGE_SHIFT;
422 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
423
424 memset(&nodes, 0, sizeof(nodes));
425
426
427
428
429 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
430 long n = simple_strtol(cmdline, NULL, 0);
431
432 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
433 if (num_nodes < 0)
434 return num_nodes;
435 goto out;
436 }
437
438
439 for (coeff_flag = 0; ; cmdline++) {
440 if (*cmdline && isdigit(*cmdline)) {
441 num = num * 10 + *cmdline - '0';
442 continue;
443 }
444 if (*cmdline == '*') {
445 if (num > 0)
446 coeff = num;
447 coeff_flag = 1;
448 }
449 if (!*cmdline || *cmdline == ',') {
450 if (!coeff_flag)
451 coeff = 1;
452
453
454
455
456 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
457 if (size)
458 for (i = 0; i < coeff; i++, num_nodes++)
459 if (setup_node_range(num_nodes, nodes,
460 &addr, size, max_addr) < 0)
461 goto done;
462 if (!*cmdline)
463 break;
464 coeff_flag = 0;
465 coeff = -1;
466 }
467 num = 0;
468 }
469done:
470 if (!num_nodes)
471 return -1;
472
473 if (addr < max_addr) {
474 if (coeff_flag && coeff < 0) {
475
476 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
477 num_nodes, num);
478 goto out;
479 }
480 switch (*(cmdline - 1)) {
481 case '*':
482
483 if (coeff <= 0)
484 break;
485 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
486 num_nodes, coeff);
487 break;
488 case ',':
489
490 break;
491 default:
492
493 setup_node_range(num_nodes, nodes, &addr,
494 max_addr - addr, max_addr);
495 num_nodes++;
496 }
497 }
498out:
499 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
500 if (memnode_shift < 0) {
501 memnode_shift = 0;
502 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
503 "disabled.\n");
504 return -1;
505 }
506
507
508
509
510
511
512 remove_all_active_ranges();
513#ifdef CONFIG_ACPI_NUMA
514 acpi_numa = -1;
515#endif
516 for_each_node_mask(i, node_possible_map) {
517 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
518 nodes[i].end >> PAGE_SHIFT);
519 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
520 }
521 acpi_fake_nodes(nodes, num_nodes);
522 numa_init_array();
523 return 0;
524}
525#endif
526
527void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn)
528{
529 int i;
530
531 nodes_clear(node_possible_map);
532 nodes_clear(node_online_map);
533
534#ifdef CONFIG_NUMA_EMU
535 if (cmdline && !numa_emulation(start_pfn, last_pfn))
536 return;
537 nodes_clear(node_possible_map);
538 nodes_clear(node_online_map);
539#endif
540
541#ifdef CONFIG_ACPI_NUMA
542 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
543 last_pfn << PAGE_SHIFT))
544 return;
545 nodes_clear(node_possible_map);
546 nodes_clear(node_online_map);
547#endif
548
549#ifdef CONFIG_K8_NUMA
550 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
551 last_pfn<<PAGE_SHIFT))
552 return;
553 nodes_clear(node_possible_map);
554 nodes_clear(node_online_map);
555#endif
556 printk(KERN_INFO "%s\n",
557 numa_off ? "NUMA turned off" : "No NUMA configuration found");
558
559 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
560 start_pfn << PAGE_SHIFT,
561 last_pfn << PAGE_SHIFT);
562
563 memnode_shift = 63;
564 memnodemap = memnode.embedded_map;
565 memnodemap[0] = 0;
566 node_set_online(0);
567 node_set(0, node_possible_map);
568 for (i = 0; i < nr_cpu_ids; i++)
569 numa_set_node(i, 0);
570 e820_register_active_regions(0, start_pfn, last_pfn);
571 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
572}
573
574unsigned long __init numa_free_all_bootmem(void)
575{
576 unsigned long pages = 0;
577 int i;
578
579 for_each_online_node(i)
580 pages += free_all_bootmem_node(NODE_DATA(i));
581
582 return pages;
583}
584
585static __init int numa_setup(char *opt)
586{
587 if (!opt)
588 return -EINVAL;
589 if (!strncmp(opt, "off", 3))
590 numa_off = 1;
591#ifdef CONFIG_NUMA_EMU
592 if (!strncmp(opt, "fake=", 5))
593 cmdline = opt + 5;
594#endif
595#ifdef CONFIG_ACPI_NUMA
596 if (!strncmp(opt, "noacpi", 6))
597 acpi_numa = -1;
598#endif
599 return 0;
600}
601early_param("numa", numa_setup);
602
603#ifdef CONFIG_NUMA
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618void __init init_cpu_to_node(void)
619{
620 int cpu;
621 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
622
623 BUG_ON(cpu_to_apicid == NULL);
624
625 for_each_possible_cpu(cpu) {
626 int node;
627 u16 apicid = cpu_to_apicid[cpu];
628
629 if (apicid == BAD_APICID)
630 continue;
631 node = apicid_to_node[apicid];
632 if (node == NUMA_NO_NODE)
633 continue;
634 if (!node_online(node))
635 continue;
636 numa_set_node(cpu, node);
637 }
638}
639#endif
640
641
642void __cpuinit numa_set_node(int cpu, int node)
643{
644 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
645
646
647 if (cpu_to_node_map) {
648 cpu_to_node_map[cpu] = node;
649 return;
650 }
651
652#ifdef CONFIG_DEBUG_PER_CPU_MAPS
653 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
654 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
655 dump_stack();
656 return;
657 }
658#endif
659 per_cpu(x86_cpu_to_node_map, cpu) = node;
660
661 if (node != NUMA_NO_NODE)
662 per_cpu(node_number, cpu) = node;
663}
664
665void __cpuinit numa_clear_node(int cpu)
666{
667 numa_set_node(cpu, NUMA_NO_NODE);
668}
669
670#ifndef CONFIG_DEBUG_PER_CPU_MAPS
671
672void __cpuinit numa_add_cpu(int cpu)
673{
674 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
675}
676
677void __cpuinit numa_remove_cpu(int cpu)
678{
679 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
680}
681
682#else
683
684
685
686
687static void __cpuinit numa_set_cpumask(int cpu, int enable)
688{
689 int node = early_cpu_to_node(cpu);
690 struct cpumask *mask;
691 char buf[64];
692
693 mask = node_to_cpumask_map[node];
694 if (mask == NULL) {
695 printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node);
696 dump_stack();
697 return;
698 }
699
700 if (enable)
701 cpumask_set_cpu(cpu, mask);
702 else
703 cpumask_clear_cpu(cpu, mask);
704
705 cpulist_scnprintf(buf, sizeof(buf), mask);
706 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
707 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
708}
709
710void __cpuinit numa_add_cpu(int cpu)
711{
712 numa_set_cpumask(cpu, 1);
713}
714
715void __cpuinit numa_remove_cpu(int cpu)
716{
717 numa_set_cpumask(cpu, 0);
718}
719
720int cpu_to_node(int cpu)
721{
722 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
723 printk(KERN_WARNING
724 "cpu_to_node(%d): usage too early!\n", cpu);
725 dump_stack();
726 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
727 }
728 return per_cpu(x86_cpu_to_node_map, cpu);
729}
730EXPORT_SYMBOL(cpu_to_node);
731
732
733
734
735
736int early_cpu_to_node(int cpu)
737{
738 if (early_per_cpu_ptr(x86_cpu_to_node_map))
739 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
740
741 if (!cpu_possible(cpu)) {
742 printk(KERN_WARNING
743 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
744 dump_stack();
745 return NUMA_NO_NODE;
746 }
747 return per_cpu(x86_cpu_to_node_map, cpu);
748}
749
750
751
752
753
754#endif
755