[PATCH] powerpc: Always panic if lmb_alloc() fails
[linux-2.6.git] / arch / powerpc / 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/sparsemem.h>
21 #include <asm/lmb.h>
22 #include <asm/system.h>
23 #include <asm/smp.h>
24
25 static int numa_enabled = 1;
26
27 static int numa_debug;
28 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
29
30 int numa_cpu_lookup_table[NR_CPUS];
31 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
32 struct pglist_data *node_data[MAX_NUMNODES];
33
34 EXPORT_SYMBOL(numa_cpu_lookup_table);
35 EXPORT_SYMBOL(numa_cpumask_lookup_table);
36 EXPORT_SYMBOL(node_data);
37
38 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
39 static int min_common_depth;
40 static int n_mem_addr_cells, n_mem_size_cells;
41
42 /*
43  * We need somewhere to store start/end/node for each region until we have
44  * allocated the real node_data structures.
45  */
46 #define MAX_REGIONS     (MAX_LMB_REGIONS*2)
47 static struct {
48         unsigned long start_pfn;
49         unsigned long end_pfn;
50         int nid;
51 } init_node_data[MAX_REGIONS] __initdata;
52
53 int __init early_pfn_to_nid(unsigned long pfn)
54 {
55         unsigned int i;
56
57         for (i = 0; init_node_data[i].end_pfn; i++) {
58                 unsigned long start_pfn = init_node_data[i].start_pfn;
59                 unsigned long end_pfn = init_node_data[i].end_pfn;
60
61                 if ((start_pfn <= pfn) && (pfn < end_pfn))
62                         return init_node_data[i].nid;
63         }
64
65         return -1;
66 }
67
68 void __init add_region(unsigned int nid, unsigned long start_pfn,
69                        unsigned long pages)
70 {
71         unsigned int i;
72
73         dbg("add_region nid %d start_pfn 0x%lx pages 0x%lx\n",
74                 nid, start_pfn, pages);
75
76         for (i = 0; init_node_data[i].end_pfn; i++) {
77                 if (init_node_data[i].nid != nid)
78                         continue;
79                 if (init_node_data[i].end_pfn == start_pfn) {
80                         init_node_data[i].end_pfn += pages;
81                         return;
82                 }
83                 if (init_node_data[i].start_pfn == (start_pfn + pages)) {
84                         init_node_data[i].start_pfn -= pages;
85                         return;
86                 }
87         }
88
89         /*
90          * Leave last entry NULL so we dont iterate off the end (we use
91          * entry.end_pfn to terminate the walk).
92          */
93         if (i >= (MAX_REGIONS - 1)) {
94                 printk(KERN_ERR "WARNING: too many memory regions in "
95                                 "numa code, truncating\n");
96                 return;
97         }
98
99         init_node_data[i].start_pfn = start_pfn;
100         init_node_data[i].end_pfn = start_pfn + pages;
101         init_node_data[i].nid = nid;
102 }
103
104 /* We assume init_node_data has no overlapping regions */
105 void __init get_region(unsigned int nid, unsigned long *start_pfn,
106                        unsigned long *end_pfn, unsigned long *pages_present)
107 {
108         unsigned int i;
109
110         *start_pfn = -1UL;
111         *end_pfn = *pages_present = 0;
112
113         for (i = 0; init_node_data[i].end_pfn; i++) {
114                 if (init_node_data[i].nid != nid)
115                         continue;
116
117                 *pages_present += init_node_data[i].end_pfn -
118                         init_node_data[i].start_pfn;
119
120                 if (init_node_data[i].start_pfn < *start_pfn)
121                         *start_pfn = init_node_data[i].start_pfn;
122
123                 if (init_node_data[i].end_pfn > *end_pfn)
124                         *end_pfn = init_node_data[i].end_pfn;
125         }
126
127         /* We didnt find a matching region, return start/end as 0 */
128         if (*start_pfn == -1UL)
129                 *start_pfn = 0;
130 }
131
132 static inline void map_cpu_to_node(int cpu, int node)
133 {
134         numa_cpu_lookup_table[cpu] = node;
135
136         if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
137                 cpu_set(cpu, numa_cpumask_lookup_table[node]);
138 }
139
140 #ifdef CONFIG_HOTPLUG_CPU
141 static void unmap_cpu_from_node(unsigned long cpu)
142 {
143         int node = numa_cpu_lookup_table[cpu];
144
145         dbg("removing cpu %lu from node %d\n", cpu, node);
146
147         if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
148                 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
149         } else {
150                 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
151                        cpu, node);
152         }
153 }
154 #endif /* CONFIG_HOTPLUG_CPU */
155
156 static struct device_node *find_cpu_node(unsigned int cpu)
157 {
158         unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
159         struct device_node *cpu_node = NULL;
160         unsigned int *interrupt_server, *reg;
161         int len;
162
163         while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
164                 /* Try interrupt server first */
165                 interrupt_server = (unsigned int *)get_property(cpu_node,
166                                         "ibm,ppc-interrupt-server#s", &len);
167
168                 len = len / sizeof(u32);
169
170                 if (interrupt_server && (len > 0)) {
171                         while (len--) {
172                                 if (interrupt_server[len] == hw_cpuid)
173                                         return cpu_node;
174                         }
175                 } else {
176                         reg = (unsigned int *)get_property(cpu_node,
177                                                            "reg", &len);
178                         if (reg && (len > 0) && (reg[0] == hw_cpuid))
179                                 return cpu_node;
180                 }
181         }
182
183         return NULL;
184 }
185
186 /* must hold reference to node during call */
187 static int *of_get_associativity(struct device_node *dev)
188 {
189         return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
190 }
191
192 static int of_node_numa_domain(struct device_node *device)
193 {
194         int numa_domain;
195         unsigned int *tmp;
196
197         if (min_common_depth == -1)
198                 return 0;
199
200         tmp = of_get_associativity(device);
201         if (tmp && (tmp[0] >= min_common_depth)) {
202                 numa_domain = tmp[min_common_depth];
203         } else {
204                 dbg("WARNING: no NUMA information for %s\n",
205                     device->full_name);
206                 numa_domain = 0;
207         }
208         return numa_domain;
209 }
210
211 /*
212  * In theory, the "ibm,associativity" property may contain multiple
213  * associativity lists because a resource may be multiply connected
214  * into the machine.  This resource then has different associativity
215  * characteristics relative to its multiple connections.  We ignore
216  * this for now.  We also assume that all cpu and memory sets have
217  * their distances represented at a common level.  This won't be
218  * true for heirarchical NUMA.
219  *
220  * In any case the ibm,associativity-reference-points should give
221  * the correct depth for a normal NUMA system.
222  *
223  * - Dave Hansen <haveblue@us.ibm.com>
224  */
225 static int __init find_min_common_depth(void)
226 {
227         int depth;
228         unsigned int *ref_points;
229         struct device_node *rtas_root;
230         unsigned int len;
231
232         rtas_root = of_find_node_by_path("/rtas");
233
234         if (!rtas_root)
235                 return -1;
236
237         /*
238          * this property is 2 32-bit integers, each representing a level of
239          * depth in the associativity nodes.  The first is for an SMP
240          * configuration (should be all 0's) and the second is for a normal
241          * NUMA configuration.
242          */
243         ref_points = (unsigned int *)get_property(rtas_root,
244                         "ibm,associativity-reference-points", &len);
245
246         if ((len >= 1) && ref_points) {
247                 depth = ref_points[1];
248         } else {
249                 dbg("WARNING: could not find NUMA "
250                     "associativity reference point\n");
251                 depth = -1;
252         }
253         of_node_put(rtas_root);
254
255         return depth;
256 }
257
258 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
259 {
260         struct device_node *memory = NULL;
261
262         memory = of_find_node_by_type(memory, "memory");
263         if (!memory)
264                 panic("numa.c: No memory nodes found!");
265
266         *n_addr_cells = prom_n_addr_cells(memory);
267         *n_size_cells = prom_n_size_cells(memory);
268         of_node_put(memory);
269 }
270
271 static unsigned long __devinit read_n_cells(int n, unsigned int **buf)
272 {
273         unsigned long result = 0;
274
275         while (n--) {
276                 result = (result << 32) | **buf;
277                 (*buf)++;
278         }
279         return result;
280 }
281
282 /*
283  * Figure out to which domain a cpu belongs and stick it there.
284  * Return the id of the domain used.
285  */
286 static int numa_setup_cpu(unsigned long lcpu)
287 {
288         int numa_domain = 0;
289         struct device_node *cpu = find_cpu_node(lcpu);
290
291         if (!cpu) {
292                 WARN_ON(1);
293                 goto out;
294         }
295
296         numa_domain = of_node_numa_domain(cpu);
297
298         if (numa_domain >= num_online_nodes()) {
299                 /*
300                  * POWER4 LPAR uses 0xffff as invalid node,
301                  * dont warn in this case.
302                  */
303                 if (numa_domain != 0xffff)
304                         printk(KERN_ERR "WARNING: cpu %ld "
305                                "maps to invalid NUMA node %d\n",
306                                lcpu, numa_domain);
307                 numa_domain = 0;
308         }
309 out:
310         node_set_online(numa_domain);
311
312         map_cpu_to_node(lcpu, numa_domain);
313
314         of_node_put(cpu);
315
316         return numa_domain;
317 }
318
319 static int cpu_numa_callback(struct notifier_block *nfb,
320                              unsigned long action,
321                              void *hcpu)
322 {
323         unsigned long lcpu = (unsigned long)hcpu;
324         int ret = NOTIFY_DONE;
325
326         switch (action) {
327         case CPU_UP_PREPARE:
328                 if (min_common_depth == -1 || !numa_enabled)
329                         map_cpu_to_node(lcpu, 0);
330                 else
331                         numa_setup_cpu(lcpu);
332                 ret = NOTIFY_OK;
333                 break;
334 #ifdef CONFIG_HOTPLUG_CPU
335         case CPU_DEAD:
336         case CPU_UP_CANCELED:
337                 unmap_cpu_from_node(lcpu);
338                 break;
339                 ret = NOTIFY_OK;
340 #endif
341         }
342         return ret;
343 }
344
345 /*
346  * Check and possibly modify a memory region to enforce the memory limit.
347  *
348  * Returns the size the region should have to enforce the memory limit.
349  * This will either be the original value of size, a truncated value,
350  * or zero. If the returned value of size is 0 the region should be
351  * discarded as it lies wholy above the memory limit.
352  */
353 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
354                                                       unsigned long size)
355 {
356         /*
357          * We use lmb_end_of_DRAM() in here instead of memory_limit because
358          * we've already adjusted it for the limit and it takes care of
359          * having memory holes below the limit.
360          */
361
362         if (! memory_limit)
363                 return size;
364
365         if (start + size <= lmb_end_of_DRAM())
366                 return size;
367
368         if (start >= lmb_end_of_DRAM())
369                 return 0;
370
371         return lmb_end_of_DRAM() - start;
372 }
373
374 static int __init parse_numa_properties(void)
375 {
376         struct device_node *cpu = NULL;
377         struct device_node *memory = NULL;
378         int max_domain;
379         unsigned long i;
380
381         if (numa_enabled == 0) {
382                 printk(KERN_WARNING "NUMA disabled by user\n");
383                 return -1;
384         }
385
386         min_common_depth = find_min_common_depth();
387
388         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
389         if (min_common_depth < 0)
390                 return min_common_depth;
391
392         max_domain = numa_setup_cpu(boot_cpuid);
393
394         /*
395          * Even though we connect cpus to numa domains later in SMP init,
396          * we need to know the maximum node id now. This is because each
397          * node id must have NODE_DATA etc backing it.
398          * As a result of hotplug we could still have cpus appear later on
399          * with larger node ids. In that case we force the cpu into node 0.
400          */
401         for_each_cpu(i) {
402                 int numa_domain;
403
404                 cpu = find_cpu_node(i);
405
406                 if (cpu) {
407                         numa_domain = of_node_numa_domain(cpu);
408                         of_node_put(cpu);
409
410                         if (numa_domain < MAX_NUMNODES &&
411                             max_domain < numa_domain)
412                                 max_domain = numa_domain;
413                 }
414         }
415
416         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
417         memory = NULL;
418         while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
419                 unsigned long start;
420                 unsigned long size;
421                 int numa_domain;
422                 int ranges;
423                 unsigned int *memcell_buf;
424                 unsigned int len;
425
426                 memcell_buf = (unsigned int *)get_property(memory,
427                         "linux,usable-memory", &len);
428                 if (!memcell_buf || len <= 0)
429                         memcell_buf =
430                                 (unsigned int *)get_property(memory, "reg",
431                                         &len);
432                 if (!memcell_buf || len <= 0)
433                         continue;
434
435                 /* ranges in cell */
436                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
437 new_range:
438                 /* these are order-sensitive, and modify the buffer pointer */
439                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
440                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
441
442                 numa_domain = of_node_numa_domain(memory);
443
444                 if (numa_domain >= MAX_NUMNODES) {
445                         if (numa_domain != 0xffff)
446                                 printk(KERN_ERR "WARNING: memory at %lx maps "
447                                        "to invalid NUMA node %d\n", start,
448                                        numa_domain);
449                         numa_domain = 0;
450                 }
451
452                 if (max_domain < numa_domain)
453                         max_domain = numa_domain;
454
455                 if (!(size = numa_enforce_memory_limit(start, size))) {
456                         if (--ranges)
457                                 goto new_range;
458                         else
459                                 continue;
460                 }
461
462                 add_region(numa_domain, start >> PAGE_SHIFT,
463                            size >> PAGE_SHIFT);
464
465                 if (--ranges)
466                         goto new_range;
467         }
468
469         for (i = 0; i <= max_domain; i++)
470                 node_set_online(i);
471
472         return 0;
473 }
474
475 static void __init setup_nonnuma(void)
476 {
477         unsigned long top_of_ram = lmb_end_of_DRAM();
478         unsigned long total_ram = lmb_phys_mem_size();
479         unsigned int i;
480
481         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
482                top_of_ram, total_ram);
483         printk(KERN_INFO "Memory hole size: %ldMB\n",
484                (top_of_ram - total_ram) >> 20);
485
486         map_cpu_to_node(boot_cpuid, 0);
487         for (i = 0; i < lmb.memory.cnt; ++i)
488                 add_region(0, lmb.memory.region[i].base >> PAGE_SHIFT,
489                            lmb_size_pages(&lmb.memory, i));
490         node_set_online(0);
491 }
492
493 void __init dump_numa_cpu_topology(void)
494 {
495         unsigned int node;
496         unsigned int cpu, count;
497
498         if (min_common_depth == -1 || !numa_enabled)
499                 return;
500
501         for_each_online_node(node) {
502                 printk(KERN_INFO "Node %d CPUs:", node);
503
504                 count = 0;
505                 /*
506                  * If we used a CPU iterator here we would miss printing
507                  * the holes in the cpumap.
508                  */
509                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
510                         if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
511                                 if (count == 0)
512                                         printk(" %u", cpu);
513                                 ++count;
514                         } else {
515                                 if (count > 1)
516                                         printk("-%u", cpu - 1);
517                                 count = 0;
518                         }
519                 }
520
521                 if (count > 1)
522                         printk("-%u", NR_CPUS - 1);
523                 printk("\n");
524         }
525 }
526
527 static void __init dump_numa_memory_topology(void)
528 {
529         unsigned int node;
530         unsigned int count;
531
532         if (min_common_depth == -1 || !numa_enabled)
533                 return;
534
535         for_each_online_node(node) {
536                 unsigned long i;
537
538                 printk(KERN_INFO "Node %d Memory:", node);
539
540                 count = 0;
541
542                 for (i = 0; i < lmb_end_of_DRAM();
543                      i += (1 << SECTION_SIZE_BITS)) {
544                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
545                                 if (count == 0)
546                                         printk(" 0x%lx", i);
547                                 ++count;
548                         } else {
549                                 if (count > 0)
550                                         printk("-0x%lx", i);
551                                 count = 0;
552                         }
553                 }
554
555                 if (count > 0)
556                         printk("-0x%lx", i);
557                 printk("\n");
558         }
559 }
560
561 /*
562  * Allocate some memory, satisfying the lmb or bootmem allocator where
563  * required. nid is the preferred node and end is the physical address of
564  * the highest address in the node.
565  *
566  * Returns the physical address of the memory.
567  */
568 static void __init *careful_allocation(int nid, unsigned long size,
569                                        unsigned long align,
570                                        unsigned long end_pfn)
571 {
572         int new_nid;
573         unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
574
575         /* retry over all memory */
576         if (!ret)
577                 ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
578
579         if (!ret)
580                 panic("numa.c: cannot allocate %lu bytes on node %d",
581                       size, nid);
582
583         /*
584          * If the memory came from a previously allocated node, we must
585          * retry with the bootmem allocator.
586          */
587         new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
588         if (new_nid < nid) {
589                 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
590                                 size, align, 0);
591
592                 if (!ret)
593                         panic("numa.c: cannot allocate %lu bytes on node %d",
594                               size, new_nid);
595
596                 ret = __pa(ret);
597
598                 dbg("alloc_bootmem %lx %lx\n", ret, size);
599         }
600
601         return (void *)ret;
602 }
603
604 void __init do_init_bootmem(void)
605 {
606         int nid;
607         unsigned int i;
608         static struct notifier_block ppc64_numa_nb = {
609                 .notifier_call = cpu_numa_callback,
610                 .priority = 1 /* Must run before sched domains notifier. */
611         };
612
613         min_low_pfn = 0;
614         max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
615         max_pfn = max_low_pfn;
616
617         if (parse_numa_properties())
618                 setup_nonnuma();
619         else
620                 dump_numa_memory_topology();
621
622         register_cpu_notifier(&ppc64_numa_nb);
623
624         for_each_online_node(nid) {
625                 unsigned long start_pfn, end_pfn, pages_present;
626                 unsigned long bootmem_paddr;
627                 unsigned long bootmap_pages;
628
629                 get_region(nid, &start_pfn, &end_pfn, &pages_present);
630
631                 /* Allocate the node structure node local if possible */
632                 NODE_DATA(nid) = careful_allocation(nid,
633                                         sizeof(struct pglist_data),
634                                         SMP_CACHE_BYTES, end_pfn);
635                 NODE_DATA(nid) = __va(NODE_DATA(nid));
636                 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
637
638                 dbg("node %d\n", nid);
639                 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
640
641                 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
642                 NODE_DATA(nid)->node_start_pfn = start_pfn;
643                 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
644
645                 if (NODE_DATA(nid)->node_spanned_pages == 0)
646                         continue;
647
648                 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
649                 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
650
651                 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
652                 bootmem_paddr = (unsigned long)careful_allocation(nid,
653                                         bootmap_pages << PAGE_SHIFT,
654                                         PAGE_SIZE, end_pfn);
655                 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
656
657                 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
658
659                 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
660                                   start_pfn, end_pfn);
661
662                 /* Add free regions on this node */
663                 for (i = 0; init_node_data[i].end_pfn; i++) {
664                         unsigned long start, end;
665
666                         if (init_node_data[i].nid != nid)
667                                 continue;
668
669                         start = init_node_data[i].start_pfn << PAGE_SHIFT;
670                         end = init_node_data[i].end_pfn << PAGE_SHIFT;
671
672                         dbg("free_bootmem %lx %lx\n", start, end - start);
673                         free_bootmem_node(NODE_DATA(nid), start, end - start);
674                 }
675
676                 /* Mark reserved regions on this node */
677                 for (i = 0; i < lmb.reserved.cnt; i++) {
678                         unsigned long physbase = lmb.reserved.region[i].base;
679                         unsigned long size = lmb.reserved.region[i].size;
680                         unsigned long start_paddr = start_pfn << PAGE_SHIFT;
681                         unsigned long end_paddr = end_pfn << PAGE_SHIFT;
682
683                         if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
684                             early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
685                                 continue;
686
687                         if (physbase < end_paddr &&
688                             (physbase+size) > start_paddr) {
689                                 /* overlaps */
690                                 if (physbase < start_paddr) {
691                                         size -= start_paddr - physbase;
692                                         physbase = start_paddr;
693                                 }
694
695                                 if (size > end_paddr - physbase)
696                                         size = end_paddr - physbase;
697
698                                 dbg("reserve_bootmem %lx %lx\n", physbase,
699                                     size);
700                                 reserve_bootmem_node(NODE_DATA(nid), physbase,
701                                                      size);
702                         }
703                 }
704
705                 /* Add regions into sparsemem */
706                 for (i = 0; init_node_data[i].end_pfn; i++) {
707                         unsigned long start, end;
708
709                         if (init_node_data[i].nid != nid)
710                                 continue;
711
712                         start = init_node_data[i].start_pfn;
713                         end = init_node_data[i].end_pfn;
714
715                         memory_present(nid, start, end);
716                 }
717         }
718 }
719
720 void __init paging_init(void)
721 {
722         unsigned long zones_size[MAX_NR_ZONES];
723         unsigned long zholes_size[MAX_NR_ZONES];
724         int nid;
725
726         memset(zones_size, 0, sizeof(zones_size));
727         memset(zholes_size, 0, sizeof(zholes_size));
728
729         for_each_online_node(nid) {
730                 unsigned long start_pfn, end_pfn, pages_present;
731
732                 get_region(nid, &start_pfn, &end_pfn, &pages_present);
733
734                 zones_size[ZONE_DMA] = end_pfn - start_pfn;
735                 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] - pages_present;
736
737                 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
738                     zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
739
740                 free_area_init_node(nid, NODE_DATA(nid), zones_size, start_pfn,
741                                     zholes_size);
742         }
743 }
744
745 static int __init early_numa(char *p)
746 {
747         if (!p)
748                 return 0;
749
750         if (strstr(p, "off"))
751                 numa_enabled = 0;
752
753         if (strstr(p, "debug"))
754                 numa_debug = 1;
755
756         return 0;
757 }
758 early_param("numa", early_numa);
759
760 #ifdef CONFIG_MEMORY_HOTPLUG
761 /*
762  * Find the node associated with a hot added memory section.  Section
763  * corresponds to a SPARSEMEM section, not an LMB.  It is assumed that
764  * sections are fully contained within a single LMB.
765  */
766 int hot_add_scn_to_nid(unsigned long scn_addr)
767 {
768         struct device_node *memory = NULL;
769         nodemask_t nodes;
770         int numa_domain = 0;
771
772         if (!numa_enabled || (min_common_depth < 0))
773                 return numa_domain;
774
775         while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
776                 unsigned long start, size;
777                 int ranges;
778                 unsigned int *memcell_buf;
779                 unsigned int len;
780
781                 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
782                 if (!memcell_buf || len <= 0)
783                         continue;
784
785                 /* ranges in cell */
786                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
787 ha_new_range:
788                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
789                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
790                 numa_domain = of_node_numa_domain(memory);
791
792                 /* Domains not present at boot default to 0 */
793                 if (!node_online(numa_domain))
794                         numa_domain = any_online_node(NODE_MASK_ALL);
795
796                 if ((scn_addr >= start) && (scn_addr < (start + size))) {
797                         of_node_put(memory);
798                         goto got_numa_domain;
799                 }
800
801                 if (--ranges)           /* process all ranges in cell */
802                         goto ha_new_range;
803         }
804         BUG();  /* section address should be found above */
805
806         /* Temporary code to ensure that returned node is not empty */
807 got_numa_domain:
808         nodes_setall(nodes);
809         while (NODE_DATA(numa_domain)->node_spanned_pages == 0) {
810                 node_clear(numa_domain, nodes);
811                 numa_domain = any_online_node(nodes);
812         }
813         return numa_domain;
814 }
815 #endif /* CONFIG_MEMORY_HOTPLUG */