blob: ea862ec643d3a9df7f6f2f5287c1c81402f1f321 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * pSeries NUMA support
3 *
4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 */
11#include <linux/threads.h>
12#include <linux/bootmem.h>
13#include <linux/init.h>
14#include <linux/mm.h>
15#include <linux/mmzone.h>
16#include <linux/module.h>
17#include <linux/nodemask.h>
18#include <linux/cpu.h>
19#include <linux/notifier.h>
20#include <asm/lmb.h>
21#include <asm/machdep.h>
22#include <asm/abs_addr.h>
23
24static int numa_enabled = 1;
25
26static int numa_debug;
27#define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
28
29#ifdef DEBUG_NUMA
30#define ARRAY_INITIALISER -1
31#else
32#define ARRAY_INITIALISER 0
33#endif
34
35int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
36 ARRAY_INITIALISER};
37char *numa_memory_lookup_table;
38cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
39int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
40
41struct pglist_data *node_data[MAX_NUMNODES];
42bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
43static int min_common_depth;
44
45/*
46 * We need somewhere to store start/span for each node until we have
47 * allocated the real node_data structures.
48 */
49static struct {
50 unsigned long node_start_pfn;
51 unsigned long node_end_pfn;
52 unsigned long node_present_pages;
53} init_node_data[MAX_NUMNODES] __initdata;
54
55EXPORT_SYMBOL(node_data);
56EXPORT_SYMBOL(numa_cpu_lookup_table);
57EXPORT_SYMBOL(numa_memory_lookup_table);
58EXPORT_SYMBOL(numa_cpumask_lookup_table);
59EXPORT_SYMBOL(nr_cpus_in_node);
60
61static inline void map_cpu_to_node(int cpu, int node)
62{
63 numa_cpu_lookup_table[cpu] = node;
64 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node]))) {
65 cpu_set(cpu, numa_cpumask_lookup_table[node]);
66 nr_cpus_in_node[node]++;
67 }
68}
69
70#ifdef CONFIG_HOTPLUG_CPU
71static void unmap_cpu_from_node(unsigned long cpu)
72{
73 int node = numa_cpu_lookup_table[cpu];
74
75 dbg("removing cpu %lu from node %d\n", cpu, node);
76
77 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
78 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
79 nr_cpus_in_node[node]--;
80 } else {
81 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
82 cpu, node);
83 }
84}
85#endif /* CONFIG_HOTPLUG_CPU */
86
87static struct device_node * __devinit find_cpu_node(unsigned int cpu)
88{
89 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
90 struct device_node *cpu_node = NULL;
91 unsigned int *interrupt_server, *reg;
92 int len;
93
94 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
95 /* Try interrupt server first */
96 interrupt_server = (unsigned int *)get_property(cpu_node,
97 "ibm,ppc-interrupt-server#s", &len);
98
99 len = len / sizeof(u32);
100
101 if (interrupt_server && (len > 0)) {
102 while (len--) {
103 if (interrupt_server[len] == hw_cpuid)
104 return cpu_node;
105 }
106 } else {
107 reg = (unsigned int *)get_property(cpu_node,
108 "reg", &len);
109 if (reg && (len > 0) && (reg[0] == hw_cpuid))
110 return cpu_node;
111 }
112 }
113
114 return NULL;
115}
116
117/* must hold reference to node during call */
118static int *of_get_associativity(struct device_node *dev)
119{
120 return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
121}
122
123static int of_node_numa_domain(struct device_node *device)
124{
125 int numa_domain;
126 unsigned int *tmp;
127
128 if (min_common_depth == -1)
129 return 0;
130
131 tmp = of_get_associativity(device);
132 if (tmp && (tmp[0] >= min_common_depth)) {
133 numa_domain = tmp[min_common_depth];
134 } else {
135 dbg("WARNING: no NUMA information for %s\n",
136 device->full_name);
137 numa_domain = 0;
138 }
139 return numa_domain;
140}
141
142/*
143 * In theory, the "ibm,associativity" property may contain multiple
144 * associativity lists because a resource may be multiply connected
145 * into the machine. This resource then has different associativity
146 * characteristics relative to its multiple connections. We ignore
147 * this for now. We also assume that all cpu and memory sets have
148 * their distances represented at a common level. This won't be
149 * true for heirarchical NUMA.
150 *
151 * In any case the ibm,associativity-reference-points should give
152 * the correct depth for a normal NUMA system.
153 *
154 * - Dave Hansen <haveblue@us.ibm.com>
155 */
156static int __init find_min_common_depth(void)
157{
158 int depth;
159 unsigned int *ref_points;
160 struct device_node *rtas_root;
161 unsigned int len;
162
163 rtas_root = of_find_node_by_path("/rtas");
164
165 if (!rtas_root)
166 return -1;
167
168 /*
169 * this property is 2 32-bit integers, each representing a level of
170 * depth in the associativity nodes. The first is for an SMP
171 * configuration (should be all 0's) and the second is for a normal
172 * NUMA configuration.
173 */
174 ref_points = (unsigned int *)get_property(rtas_root,
175 "ibm,associativity-reference-points", &len);
176
177 if ((len >= 1) && ref_points) {
178 depth = ref_points[1];
179 } else {
180 dbg("WARNING: could not find NUMA "
181 "associativity reference point\n");
182 depth = -1;
183 }
184 of_node_put(rtas_root);
185
186 return depth;
187}
188
189static int __init get_mem_addr_cells(void)
190{
191 struct device_node *memory = NULL;
192 int rc;
193
194 memory = of_find_node_by_type(memory, "memory");
195 if (!memory)
196 return 0; /* it won't matter */
197
198 rc = prom_n_addr_cells(memory);
199 return rc;
200}
201
202static int __init get_mem_size_cells(void)
203{
204 struct device_node *memory = NULL;
205 int rc;
206
207 memory = of_find_node_by_type(memory, "memory");
208 if (!memory)
209 return 0; /* it won't matter */
210 rc = prom_n_size_cells(memory);
211 return rc;
212}
213
214static unsigned long read_n_cells(int n, unsigned int **buf)
215{
216 unsigned long result = 0;
217
218 while (n--) {
219 result = (result << 32) | **buf;
220 (*buf)++;
221 }
222 return result;
223}
224
225/*
226 * Figure out to which domain a cpu belongs and stick it there.
227 * Return the id of the domain used.
228 */
229static int numa_setup_cpu(unsigned long lcpu)
230{
231 int numa_domain = 0;
232 struct device_node *cpu = find_cpu_node(lcpu);
233
234 if (!cpu) {
235 WARN_ON(1);
236 goto out;
237 }
238
239 numa_domain = of_node_numa_domain(cpu);
240
241 if (numa_domain >= num_online_nodes()) {
242 /*
243 * POWER4 LPAR uses 0xffff as invalid node,
244 * dont warn in this case.
245 */
246 if (numa_domain != 0xffff)
247 printk(KERN_ERR "WARNING: cpu %ld "
248 "maps to invalid NUMA node %d\n",
249 lcpu, numa_domain);
250 numa_domain = 0;
251 }
252out:
253 node_set_online(numa_domain);
254
255 map_cpu_to_node(lcpu, numa_domain);
256
257 of_node_put(cpu);
258
259 return numa_domain;
260}
261
262static int cpu_numa_callback(struct notifier_block *nfb,
263 unsigned long action,
264 void *hcpu)
265{
266 unsigned long lcpu = (unsigned long)hcpu;
267 int ret = NOTIFY_DONE;
268
269 switch (action) {
270 case CPU_UP_PREPARE:
271 if (min_common_depth == -1 || !numa_enabled)
272 map_cpu_to_node(lcpu, 0);
273 else
274 numa_setup_cpu(lcpu);
275 ret = NOTIFY_OK;
276 break;
277#ifdef CONFIG_HOTPLUG_CPU
278 case CPU_DEAD:
279 case CPU_UP_CANCELED:
280 unmap_cpu_from_node(lcpu);
281 break;
282 ret = NOTIFY_OK;
283#endif
284 }
285 return ret;
286}
287
288/*
289 * Check and possibly modify a memory region to enforce the memory limit.
290 *
291 * Returns the size the region should have to enforce the memory limit.
292 * This will either be the original value of size, a truncated value,
293 * or zero. If the returned value of size is 0 the region should be
294 * discarded as it lies wholy above the memory limit.
295 */
296static unsigned long __init numa_enforce_memory_limit(unsigned long start, unsigned long size)
297{
298 /*
299 * We use lmb_end_of_DRAM() in here instead of memory_limit because
300 * we've already adjusted it for the limit and it takes care of
301 * having memory holes below the limit.
302 */
303 extern unsigned long memory_limit;
304
305 if (! memory_limit)
306 return size;
307
308 if (start + size <= lmb_end_of_DRAM())
309 return size;
310
311 if (start >= lmb_end_of_DRAM())
312 return 0;
313
314 return lmb_end_of_DRAM() - start;
315}
316
317static int __init parse_numa_properties(void)
318{
319 struct device_node *cpu = NULL;
320 struct device_node *memory = NULL;
321 int addr_cells, size_cells;
322 int max_domain = 0;
323 long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
324 unsigned long i;
325
326 if (numa_enabled == 0) {
327 printk(KERN_WARNING "NUMA disabled by user\n");
328 return -1;
329 }
330
331 numa_memory_lookup_table =
332 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
333 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
334
335 for (i = 0; i < entries ; i++)
336 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
337
338 min_common_depth = find_min_common_depth();
339
340 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
341 if (min_common_depth < 0)
342 return min_common_depth;
343
344 max_domain = numa_setup_cpu(boot_cpuid);
345
346 /*
347 * Even though we connect cpus to numa domains later in SMP init,
348 * we need to know the maximum node id now. This is because each
349 * node id must have NODE_DATA etc backing it.
350 * As a result of hotplug we could still have cpus appear later on
351 * with larger node ids. In that case we force the cpu into node 0.
352 */
353 for_each_cpu(i) {
354 int numa_domain;
355
356 cpu = find_cpu_node(i);
357
358 if (cpu) {
359 numa_domain = of_node_numa_domain(cpu);
360 of_node_put(cpu);
361
362 if (numa_domain < MAX_NUMNODES &&
363 max_domain < numa_domain)
364 max_domain = numa_domain;
365 }
366 }
367
368 addr_cells = get_mem_addr_cells();
369 size_cells = get_mem_size_cells();
370 memory = NULL;
371 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
372 unsigned long start;
373 unsigned long size;
374 int numa_domain;
375 int ranges;
376 unsigned int *memcell_buf;
377 unsigned int len;
378
379 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
380 if (!memcell_buf || len <= 0)
381 continue;
382
383 ranges = memory->n_addrs;
384new_range:
385 /* these are order-sensitive, and modify the buffer pointer */
386 start = read_n_cells(addr_cells, &memcell_buf);
387 size = read_n_cells(size_cells, &memcell_buf);
388
389 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
390 size = _ALIGN_UP(size, MEMORY_INCREMENT);
391
392 numa_domain = of_node_numa_domain(memory);
393
394 if (numa_domain >= MAX_NUMNODES) {
395 if (numa_domain != 0xffff)
396 printk(KERN_ERR "WARNING: memory at %lx maps "
397 "to invalid NUMA node %d\n", start,
398 numa_domain);
399 numa_domain = 0;
400 }
401
402 if (max_domain < numa_domain)
403 max_domain = numa_domain;
404
405 if (! (size = numa_enforce_memory_limit(start, size))) {
406 if (--ranges)
407 goto new_range;
408 else
409 continue;
410 }
411
412 /*
413 * Initialize new node struct, or add to an existing one.
414 */
415 if (init_node_data[numa_domain].node_end_pfn) {
416 if ((start / PAGE_SIZE) <
417 init_node_data[numa_domain].node_start_pfn)
418 init_node_data[numa_domain].node_start_pfn =
419 start / PAGE_SIZE;
420 if (((start / PAGE_SIZE) + (size / PAGE_SIZE)) >
421 init_node_data[numa_domain].node_end_pfn)
422 init_node_data[numa_domain].node_end_pfn =
423 (start / PAGE_SIZE) +
424 (size / PAGE_SIZE);
425
426 init_node_data[numa_domain].node_present_pages +=
427 size / PAGE_SIZE;
428 } else {
429 node_set_online(numa_domain);
430
431 init_node_data[numa_domain].node_start_pfn =
432 start / PAGE_SIZE;
433 init_node_data[numa_domain].node_end_pfn =
434 init_node_data[numa_domain].node_start_pfn +
435 size / PAGE_SIZE;
436 init_node_data[numa_domain].node_present_pages =
437 size / PAGE_SIZE;
438 }
439
440 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
441 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
442 numa_domain;
443
444 if (--ranges)
445 goto new_range;
446 }
447
448 for (i = 0; i <= max_domain; i++)
449 node_set_online(i);
450
451 return 0;
452}
453
454static void __init setup_nonnuma(void)
455{
456 unsigned long top_of_ram = lmb_end_of_DRAM();
457 unsigned long total_ram = lmb_phys_mem_size();
458 unsigned long i;
459
460 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
461 top_of_ram, total_ram);
462 printk(KERN_INFO "Memory hole size: %ldMB\n",
463 (top_of_ram - total_ram) >> 20);
464
465 if (!numa_memory_lookup_table) {
466 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
467 numa_memory_lookup_table =
468 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
469 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
470 for (i = 0; i < entries ; i++)
471 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
472 }
473
474 map_cpu_to_node(boot_cpuid, 0);
475
476 node_set_online(0);
477
478 init_node_data[0].node_start_pfn = 0;
479 init_node_data[0].node_end_pfn = lmb_end_of_DRAM() / PAGE_SIZE;
480 init_node_data[0].node_present_pages = total_ram / PAGE_SIZE;
481
482 for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
483 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
484}
485
486static void __init dump_numa_topology(void)
487{
488 unsigned int node;
489 unsigned int count;
490
491 if (min_common_depth == -1 || !numa_enabled)
492 return;
493
494 for_each_online_node(node) {
495 unsigned long i;
496
497 printk(KERN_INFO "Node %d Memory:", node);
498
499 count = 0;
500
501 for (i = 0; i < lmb_end_of_DRAM(); i += MEMORY_INCREMENT) {
502 if (numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] == node) {
503 if (count == 0)
504 printk(" 0x%lx", i);
505 ++count;
506 } else {
507 if (count > 0)
508 printk("-0x%lx", i);
509 count = 0;
510 }
511 }
512
513 if (count > 0)
514 printk("-0x%lx", i);
515 printk("\n");
516 }
517 return;
518}
519
520/*
521 * Allocate some memory, satisfying the lmb or bootmem allocator where
522 * required. nid is the preferred node and end is the physical address of
523 * the highest address in the node.
524 *
525 * Returns the physical address of the memory.
526 */
527static unsigned long careful_allocation(int nid, unsigned long size,
528 unsigned long align, unsigned long end)
529{
530 unsigned long ret = lmb_alloc_base(size, align, end);
531
532 /* retry over all memory */
533 if (!ret)
534 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
535
536 if (!ret)
537 panic("numa.c: cannot allocate %lu bytes on node %d",
538 size, nid);
539
540 /*
541 * If the memory came from a previously allocated node, we must
542 * retry with the bootmem allocator.
543 */
544 if (pa_to_nid(ret) < nid) {
545 nid = pa_to_nid(ret);
546 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(nid),
547 size, align, 0);
548
549 if (!ret)
550 panic("numa.c: cannot allocate %lu bytes on node %d",
551 size, nid);
552
553 ret = virt_to_abs(ret);
554
555 dbg("alloc_bootmem %lx %lx\n", ret, size);
556 }
557
558 return ret;
559}
560
561void __init do_init_bootmem(void)
562{
563 int nid;
564 int addr_cells, size_cells;
565 struct device_node *memory = NULL;
566 static struct notifier_block ppc64_numa_nb = {
567 .notifier_call = cpu_numa_callback,
568 .priority = 1 /* Must run before sched domains notifier. */
569 };
570
571 min_low_pfn = 0;
572 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
573 max_pfn = max_low_pfn;
574
575 if (parse_numa_properties())
576 setup_nonnuma();
577 else
578 dump_numa_topology();
579
580 register_cpu_notifier(&ppc64_numa_nb);
581
582 for_each_online_node(nid) {
583 unsigned long start_paddr, end_paddr;
584 int i;
585 unsigned long bootmem_paddr;
586 unsigned long bootmap_pages;
587
588 start_paddr = init_node_data[nid].node_start_pfn * PAGE_SIZE;
589 end_paddr = init_node_data[nid].node_end_pfn * PAGE_SIZE;
590
591 /* Allocate the node structure node local if possible */
592 NODE_DATA(nid) = (struct pglist_data *)careful_allocation(nid,
593 sizeof(struct pglist_data),
594 SMP_CACHE_BYTES, end_paddr);
595 NODE_DATA(nid) = abs_to_virt(NODE_DATA(nid));
596 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
597
598 dbg("node %d\n", nid);
599 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
600
601 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
602 NODE_DATA(nid)->node_start_pfn =
603 init_node_data[nid].node_start_pfn;
604 NODE_DATA(nid)->node_spanned_pages =
605 end_paddr - start_paddr;
606
607 if (NODE_DATA(nid)->node_spanned_pages == 0)
608 continue;
609
610 dbg("start_paddr = %lx\n", start_paddr);
611 dbg("end_paddr = %lx\n", end_paddr);
612
613 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
614
615 bootmem_paddr = careful_allocation(nid,
616 bootmap_pages << PAGE_SHIFT,
617 PAGE_SIZE, end_paddr);
618 memset(abs_to_virt(bootmem_paddr), 0,
619 bootmap_pages << PAGE_SHIFT);
620 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
621
622 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
623 start_paddr >> PAGE_SHIFT,
624 end_paddr >> PAGE_SHIFT);
625
626 /*
627 * We need to do another scan of all memory sections to
628 * associate memory with the correct node.
629 */
630 addr_cells = get_mem_addr_cells();
631 size_cells = get_mem_size_cells();
632 memory = NULL;
633 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
634 unsigned long mem_start, mem_size;
635 int numa_domain, ranges;
636 unsigned int *memcell_buf;
637 unsigned int len;
638
639 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
640 if (!memcell_buf || len <= 0)
641 continue;
642
643 ranges = memory->n_addrs; /* ranges in cell */
644new_range:
645 mem_start = read_n_cells(addr_cells, &memcell_buf);
646 mem_size = read_n_cells(size_cells, &memcell_buf);
647 numa_domain = numa_enabled ? of_node_numa_domain(memory) : 0;
648
649 if (numa_domain != nid)
650 continue;
651
652 mem_size = numa_enforce_memory_limit(mem_start, mem_size);
653 if (mem_size) {
654 dbg("free_bootmem %lx %lx\n", mem_start, mem_size);
655 free_bootmem_node(NODE_DATA(nid), mem_start, mem_size);
656 }
657
658 if (--ranges) /* process all ranges in cell */
659 goto new_range;
660 }
661
662 /*
663 * Mark reserved regions on this node
664 */
665 for (i = 0; i < lmb.reserved.cnt; i++) {
666 unsigned long physbase = lmb.reserved.region[i].physbase;
667 unsigned long size = lmb.reserved.region[i].size;
668
669 if (pa_to_nid(physbase) != nid &&
670 pa_to_nid(physbase+size-1) != nid)
671 continue;
672
673 if (physbase < end_paddr &&
674 (physbase+size) > start_paddr) {
675 /* overlaps */
676 if (physbase < start_paddr) {
677 size -= start_paddr - physbase;
678 physbase = start_paddr;
679 }
680
681 if (size > end_paddr - physbase)
682 size = end_paddr - physbase;
683
684 dbg("reserve_bootmem %lx %lx\n", physbase,
685 size);
686 reserve_bootmem_node(NODE_DATA(nid), physbase,
687 size);
688 }
689 }
690 }
691}
692
693void __init paging_init(void)
694{
695 unsigned long zones_size[MAX_NR_ZONES];
696 unsigned long zholes_size[MAX_NR_ZONES];
697 int nid;
698
699 memset(zones_size, 0, sizeof(zones_size));
700 memset(zholes_size, 0, sizeof(zholes_size));
701
702 for_each_online_node(nid) {
703 unsigned long start_pfn;
704 unsigned long end_pfn;
705
706 start_pfn = init_node_data[nid].node_start_pfn;
707 end_pfn = init_node_data[nid].node_end_pfn;
708
709 zones_size[ZONE_DMA] = end_pfn - start_pfn;
710 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
711 init_node_data[nid].node_present_pages;
712
713 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
714 zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
715
716 free_area_init_node(nid, NODE_DATA(nid), zones_size,
717 start_pfn, zholes_size);
718 }
719}
720
721static int __init early_numa(char *p)
722{
723 if (!p)
724 return 0;
725
726 if (strstr(p, "off"))
727 numa_enabled = 0;
728
729 if (strstr(p, "debug"))
730 numa_debug = 1;
731
732 return 0;
733}
734early_param("numa", early_numa);