vserver 1.9.5.x5
[linux-2.6.git] / arch / ppc64 / mm / numa.c
1 /*
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
24 static int numa_enabled = 1;
25
26 static 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
35 int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
36         ARRAY_INITIALISER};
37 char *numa_memory_lookup_table;
38 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
39 int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
40
41 struct pglist_data *node_data[MAX_NUMNODES];
42 bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
43 static unsigned long node0_io_hole_size;
44 static int min_common_depth;
45
46 /*
47  * We need somewhere to store start/span for each node until we have
48  * allocated the real node_data structures.
49  */
50 static struct {
51         unsigned long node_start_pfn;
52         unsigned long node_spanned_pages;
53 } init_node_data[MAX_NUMNODES] __initdata;
54
55 EXPORT_SYMBOL(node_data);
56 EXPORT_SYMBOL(numa_cpu_lookup_table);
57 EXPORT_SYMBOL(numa_memory_lookup_table);
58 EXPORT_SYMBOL(numa_cpumask_lookup_table);
59 EXPORT_SYMBOL(nr_cpus_in_node);
60
61 static 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
71 static 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
87 static 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 */
118 static int *of_get_associativity(struct device_node *dev)
119 {
120         return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
121 }
122
123 static 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  */
156 static 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
189 static unsigned long read_cell_ul(struct device_node *device, unsigned int **buf)
190 {
191         int i;
192         unsigned long result = 0;
193
194         i = prom_n_size_cells(device);
195         /* bug on i>2 ?? */
196         while (i--) {
197                 result = (result << 32) | **buf;
198                 (*buf)++;
199         }
200         return result;
201 }
202
203 /*
204  * Figure out to which domain a cpu belongs and stick it there.
205  * Return the id of the domain used.
206  */
207 static int numa_setup_cpu(unsigned long lcpu)
208 {
209         int numa_domain = 0;
210         struct device_node *cpu = find_cpu_node(lcpu);
211
212         if (!cpu) {
213                 WARN_ON(1);
214                 goto out;
215         }
216
217         numa_domain = of_node_numa_domain(cpu);
218
219         if (numa_domain >= num_online_nodes()) {
220                 /*
221                  * POWER4 LPAR uses 0xffff as invalid node,
222                  * dont warn in this case.
223                  */
224                 if (numa_domain != 0xffff)
225                         printk(KERN_ERR "WARNING: cpu %ld "
226                                "maps to invalid NUMA node %d\n",
227                                lcpu, numa_domain);
228                 numa_domain = 0;
229         }
230 out:
231         node_set_online(numa_domain);
232
233         map_cpu_to_node(lcpu, numa_domain);
234
235         of_node_put(cpu);
236
237         return numa_domain;
238 }
239
240 static int cpu_numa_callback(struct notifier_block *nfb,
241                              unsigned long action,
242                              void *hcpu)
243 {
244         unsigned long lcpu = (unsigned long)hcpu;
245         int ret = NOTIFY_DONE;
246
247         switch (action) {
248         case CPU_UP_PREPARE:
249                 if (min_common_depth == -1 || !numa_enabled)
250                         map_cpu_to_node(lcpu, 0);
251                 else
252                         numa_setup_cpu(lcpu);
253                 ret = NOTIFY_OK;
254                 break;
255 #ifdef CONFIG_HOTPLUG_CPU
256         case CPU_DEAD:
257         case CPU_UP_CANCELED:
258                 unmap_cpu_from_node(lcpu);
259                 break;
260                 ret = NOTIFY_OK;
261 #endif
262         }
263         return ret;
264 }
265
266 static int __init parse_numa_properties(void)
267 {
268         struct device_node *cpu = NULL;
269         struct device_node *memory = NULL;
270         int max_domain = 0;
271         long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
272         unsigned long i;
273
274         if (numa_enabled == 0) {
275                 printk(KERN_WARNING "NUMA disabled by user\n");
276                 return -1;
277         }
278
279         numa_memory_lookup_table =
280                 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
281         memset(numa_memory_lookup_table, 0, entries * sizeof(char));
282
283         for (i = 0; i < entries ; i++)
284                 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
285
286         min_common_depth = find_min_common_depth();
287
288         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
289         if (min_common_depth < 0)
290                 return min_common_depth;
291
292         max_domain = numa_setup_cpu(boot_cpuid);
293
294         /*
295          * Even though we connect cpus to numa domains later in SMP init,
296          * we need to know the maximum node id now. This is because each
297          * node id must have NODE_DATA etc backing it.
298          * As a result of hotplug we could still have cpus appear later on
299          * with larger node ids. In that case we force the cpu into node 0.
300          */
301         for_each_cpu(i) {
302                 int numa_domain;
303
304                 cpu = find_cpu_node(i);
305
306                 if (cpu) {
307                         numa_domain = of_node_numa_domain(cpu);
308                         of_node_put(cpu);
309
310                         if (numa_domain < MAX_NUMNODES &&
311                             max_domain < numa_domain)
312                                 max_domain = numa_domain;
313                 }
314         }
315
316         memory = NULL;
317         while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
318                 unsigned long start;
319                 unsigned long size;
320                 int numa_domain;
321                 int ranges;
322                 unsigned int *memcell_buf;
323                 unsigned int len;
324
325                 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
326                 if (!memcell_buf || len <= 0)
327                         continue;
328
329                 ranges = memory->n_addrs;
330 new_range:
331                 /* these are order-sensitive, and modify the buffer pointer */
332                 start = read_cell_ul(memory, &memcell_buf);
333                 size = read_cell_ul(memory, &memcell_buf);
334
335                 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
336                 size = _ALIGN_UP(size, MEMORY_INCREMENT);
337
338                 numa_domain = of_node_numa_domain(memory);
339
340                 if (numa_domain >= MAX_NUMNODES) {
341                         if (numa_domain != 0xffff)
342                                 printk(KERN_ERR "WARNING: memory at %lx maps "
343                                        "to invalid NUMA node %d\n", start,
344                                        numa_domain);
345                         numa_domain = 0;
346                 }
347
348                 if (max_domain < numa_domain)
349                         max_domain = numa_domain;
350
351                 /* 
352                  * For backwards compatibility, OF splits the first node
353                  * into two regions (the first being 0-4GB). Check for
354                  * this simple case and complain if there is a gap in
355                  * memory
356                  */
357                 if (init_node_data[numa_domain].node_spanned_pages) {
358                         unsigned long shouldstart =
359                                 init_node_data[numa_domain].node_start_pfn +
360                                 init_node_data[numa_domain].node_spanned_pages;
361                         if (shouldstart != (start / PAGE_SIZE)) {
362                                 /* Revert to non-numa for now */
363                                 printk(KERN_ERR
364                                        "WARNING: Unexpected node layout: "
365                                        "region start %lx length %lx\n",
366                                        start, size);
367                                 printk(KERN_ERR "NUMA is disabled\n");
368                                 goto err;
369                         }
370                         init_node_data[numa_domain].node_spanned_pages +=
371                                 size / PAGE_SIZE;
372                 } else {
373                         node_set_online(numa_domain);
374
375                         init_node_data[numa_domain].node_start_pfn =
376                                 start / PAGE_SIZE;
377                         init_node_data[numa_domain].node_spanned_pages =
378                                 size / PAGE_SIZE;
379                 }
380
381                 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
382                         numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
383                                 numa_domain;
384
385                 ranges--;
386                 if (ranges)
387                         goto new_range;
388         }
389
390         for (i = 0; i <= max_domain; i++)
391                 node_set_online(i);
392
393         return 0;
394 err:
395         /* Something has gone wrong; revert any setup we've done */
396         for_each_node(i) {
397                 node_set_offline(i);
398                 init_node_data[i].node_start_pfn = 0;
399                 init_node_data[i].node_spanned_pages = 0;
400         }
401         return -1;
402 }
403
404 static void __init setup_nonnuma(void)
405 {
406         unsigned long top_of_ram = lmb_end_of_DRAM();
407         unsigned long total_ram = lmb_phys_mem_size();
408         unsigned long i;
409
410         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
411                top_of_ram, total_ram);
412         printk(KERN_INFO "Memory hole size: %ldMB\n",
413                (top_of_ram - total_ram) >> 20);
414
415         if (!numa_memory_lookup_table) {
416                 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
417                 numa_memory_lookup_table =
418                         (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
419                 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
420                 for (i = 0; i < entries ; i++)
421                         numa_memory_lookup_table[i] = ARRAY_INITIALISER;
422         }
423
424         map_cpu_to_node(boot_cpuid, 0);
425
426         node_set_online(0);
427
428         init_node_data[0].node_start_pfn = 0;
429         init_node_data[0].node_spanned_pages = lmb_end_of_DRAM() / PAGE_SIZE;
430
431         for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
432                 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
433
434         node0_io_hole_size = top_of_ram - total_ram;
435 }
436
437 static void __init dump_numa_topology(void)
438 {
439         unsigned int node;
440         unsigned int count;
441
442         if (min_common_depth == -1 || !numa_enabled)
443                 return;
444
445         for_each_online_node(node) {
446                 unsigned long i;
447
448                 printk(KERN_INFO "Node %d Memory:", node);
449
450                 count = 0;
451
452                 for (i = 0; i < lmb_end_of_DRAM(); i += MEMORY_INCREMENT) {
453                         if (numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] == node) {
454                                 if (count == 0)
455                                         printk(" 0x%lx", i);
456                                 ++count;
457                         } else {
458                                 if (count > 0)
459                                         printk("-0x%lx", i);
460                                 count = 0;
461                         }
462                 }
463
464                 if (count > 0)
465                         printk("-0x%lx", i);
466                 printk("\n");
467         }
468         return;
469 }
470
471 /*
472  * Allocate some memory, satisfying the lmb or bootmem allocator where
473  * required. nid is the preferred node and end is the physical address of
474  * the highest address in the node.
475  *
476  * Returns the physical address of the memory.
477  */
478 static unsigned long careful_allocation(int nid, unsigned long size,
479                                         unsigned long align, unsigned long end)
480 {
481         unsigned long ret = lmb_alloc_base(size, align, end);
482
483         /* retry over all memory */
484         if (!ret)
485                 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
486
487         if (!ret)
488                 panic("numa.c: cannot allocate %lu bytes on node %d",
489                       size, nid);
490
491         /*
492          * If the memory came from a previously allocated node, we must
493          * retry with the bootmem allocator.
494          */
495         if (pa_to_nid(ret) < nid) {
496                 nid = pa_to_nid(ret);
497                 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(nid),
498                                 size, align, 0);
499
500                 if (!ret)
501                         panic("numa.c: cannot allocate %lu bytes on node %d",
502                               size, nid);
503
504                 ret = virt_to_abs(ret);
505
506                 dbg("alloc_bootmem %lx %lx\n", ret, size);
507         }
508
509         return ret;
510 }
511
512 void __init do_init_bootmem(void)
513 {
514         int nid;
515         static struct notifier_block ppc64_numa_nb = {
516                 .notifier_call = cpu_numa_callback,
517                 .priority = 1 /* Must run before sched domains notifier. */
518         };
519
520         min_low_pfn = 0;
521         max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
522         max_pfn = max_low_pfn;
523
524         if (parse_numa_properties())
525                 setup_nonnuma();
526         else
527                 dump_numa_topology();
528
529         register_cpu_notifier(&ppc64_numa_nb);
530
531         for_each_online_node(nid) {
532                 unsigned long start_paddr, end_paddr;
533                 int i;
534                 unsigned long bootmem_paddr;
535                 unsigned long bootmap_pages;
536
537                 start_paddr = init_node_data[nid].node_start_pfn * PAGE_SIZE;
538                 end_paddr = start_paddr + (init_node_data[nid].node_spanned_pages * PAGE_SIZE);
539
540                 /* Allocate the node structure node local if possible */
541                 NODE_DATA(nid) = (struct pglist_data *)careful_allocation(nid,
542                                         sizeof(struct pglist_data),
543                                         SMP_CACHE_BYTES, end_paddr);
544                 NODE_DATA(nid) = abs_to_virt(NODE_DATA(nid));
545                 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
546
547                 dbg("node %d\n", nid);
548                 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
549
550                 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
551                 NODE_DATA(nid)->node_start_pfn =
552                         init_node_data[nid].node_start_pfn;
553                 NODE_DATA(nid)->node_spanned_pages =
554                         init_node_data[nid].node_spanned_pages;
555
556                 if (init_node_data[nid].node_spanned_pages == 0)
557                         continue;
558
559                 dbg("start_paddr = %lx\n", start_paddr);
560                 dbg("end_paddr = %lx\n", end_paddr);
561
562                 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
563
564                 bootmem_paddr = careful_allocation(nid,
565                                 bootmap_pages << PAGE_SHIFT,
566                                 PAGE_SIZE, end_paddr);
567                 memset(abs_to_virt(bootmem_paddr), 0,
568                        bootmap_pages << PAGE_SHIFT);
569                 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
570
571                 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
572                                   start_paddr >> PAGE_SHIFT,
573                                   end_paddr >> PAGE_SHIFT);
574
575                 for (i = 0; i < lmb.memory.cnt; i++) {
576                         unsigned long physbase, size;
577
578                         physbase = lmb.memory.region[i].physbase;
579                         size = lmb.memory.region[i].size;
580
581                         if (physbase < end_paddr &&
582                             (physbase+size) > start_paddr) {
583                                 /* overlaps */
584                                 if (physbase < start_paddr) {
585                                         size -= start_paddr - physbase;
586                                         physbase = start_paddr;
587                                 }
588
589                                 if (size > end_paddr - physbase)
590                                         size = end_paddr - physbase;
591
592                                 dbg("free_bootmem %lx %lx\n", physbase, size);
593                                 free_bootmem_node(NODE_DATA(nid), physbase,
594                                                   size);
595                         }
596                 }
597
598                 for (i = 0; i < lmb.reserved.cnt; i++) {
599                         unsigned long physbase = lmb.reserved.region[i].physbase;
600                         unsigned long size = lmb.reserved.region[i].size;
601
602                         if (physbase < end_paddr &&
603                             (physbase+size) > start_paddr) {
604                                 /* overlaps */
605                                 if (physbase < start_paddr) {
606                                         size -= start_paddr - physbase;
607                                         physbase = start_paddr;
608                                 }
609
610                                 if (size > end_paddr - physbase)
611                                         size = end_paddr - physbase;
612
613                                 dbg("reserve_bootmem %lx %lx\n", physbase,
614                                     size);
615                                 reserve_bootmem_node(NODE_DATA(nid), physbase,
616                                                      size);
617                         }
618                 }
619         }
620 }
621
622 void __init paging_init(void)
623 {
624         unsigned long zones_size[MAX_NR_ZONES];
625         unsigned long zholes_size[MAX_NR_ZONES];
626         int nid;
627
628         memset(zones_size, 0, sizeof(zones_size));
629         memset(zholes_size, 0, sizeof(zholes_size));
630
631         for_each_online_node(nid) {
632                 unsigned long start_pfn;
633                 unsigned long end_pfn;
634
635                 start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
636                 end_pfn = plat_node_bdata[nid].node_low_pfn;
637
638                 zones_size[ZONE_DMA] = end_pfn - start_pfn;
639                 zholes_size[ZONE_DMA] = 0;
640                 if (nid == 0)
641                         zholes_size[ZONE_DMA] = node0_io_hole_size >> PAGE_SHIFT;
642
643                 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
644                     zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
645
646                 free_area_init_node(nid, NODE_DATA(nid), zones_size,
647                                                         start_pfn, zholes_size);
648         }
649 }
650
651 static int __init early_numa(char *p)
652 {
653         if (!p)
654                 return 0;
655
656         if (strstr(p, "off"))
657                 numa_enabled = 0;
658
659         if (strstr(p, "debug"))
660                 numa_debug = 1;
661
662         return 0;
663 }
664 early_param("numa", early_numa);