Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * pSeries NUMA support
4 *
5 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
6 */
7#define pr_fmt(fmt) "numa: " fmt
8
9#include <linux/threads.h>
10#include <linux/memblock.h>
11#include <linux/init.h>
12#include <linux/mm.h>
13#include <linux/mmzone.h>
14#include <linux/export.h>
15#include <linux/nodemask.h>
16#include <linux/cpu.h>
17#include <linux/notifier.h>
18#include <linux/of.h>
19#include <linux/pfn.h>
20#include <linux/cpuset.h>
21#include <linux/node.h>
22#include <linux/stop_machine.h>
23#include <linux/proc_fs.h>
24#include <linux/seq_file.h>
25#include <linux/uaccess.h>
26#include <linux/slab.h>
27#include <asm/cputhreads.h>
28#include <asm/sparsemem.h>
29#include <asm/smp.h>
30#include <asm/topology.h>
31#include <asm/firmware.h>
32#include <asm/paca.h>
33#include <asm/hvcall.h>
34#include <asm/setup.h>
35#include <asm/vdso.h>
36#include <asm/drmem.h>
37
38static int numa_enabled = 1;
39
40static char *cmdline __initdata;
41
42int numa_cpu_lookup_table[NR_CPUS];
43cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
44struct pglist_data *node_data[MAX_NUMNODES];
45
46EXPORT_SYMBOL(numa_cpu_lookup_table);
47EXPORT_SYMBOL(node_to_cpumask_map);
48EXPORT_SYMBOL(node_data);
49
50static int primary_domain_index;
51static int n_mem_addr_cells, n_mem_size_cells;
52
53#define FORM0_AFFINITY 0
54#define FORM1_AFFINITY 1
55#define FORM2_AFFINITY 2
56static int affinity_form;
57
58#define MAX_DISTANCE_REF_POINTS 4
59static int distance_ref_points_depth;
60static const __be32 *distance_ref_points;
61static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
62static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
63 [0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
64};
65static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
66
67/*
68 * Allocate node_to_cpumask_map based on number of available nodes
69 * Requires node_possible_map to be valid.
70 *
71 * Note: cpumask_of_node() is not valid until after this is done.
72 */
73static void __init setup_node_to_cpumask_map(void)
74{
75 unsigned int node;
76
77 /* setup nr_node_ids if not done yet */
78 if (nr_node_ids == MAX_NUMNODES)
79 setup_nr_node_ids();
80
81 /* allocate the map */
82 for_each_node(node)
83 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
84
85 /* cpumask_of_node() will now work */
86 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
87}
88
89static int __init fake_numa_create_new_node(unsigned long end_pfn,
90 unsigned int *nid)
91{
92 unsigned long long mem;
93 char *p = cmdline;
94 static unsigned int fake_nid;
95 static unsigned long long curr_boundary;
96
97 /*
98 * Modify node id, iff we started creating NUMA nodes
99 * We want to continue from where we left of the last time
100 */
101 if (fake_nid)
102 *nid = fake_nid;
103 /*
104 * In case there are no more arguments to parse, the
105 * node_id should be the same as the last fake node id
106 * (we've handled this above).
107 */
108 if (!p)
109 return 0;
110
111 mem = memparse(p, &p);
112 if (!mem)
113 return 0;
114
115 if (mem < curr_boundary)
116 return 0;
117
118 curr_boundary = mem;
119
120 if ((end_pfn << PAGE_SHIFT) > mem) {
121 /*
122 * Skip commas and spaces
123 */
124 while (*p == ',' || *p == ' ' || *p == '\t')
125 p++;
126
127 cmdline = p;
128 fake_nid++;
129 *nid = fake_nid;
130 pr_debug("created new fake_node with id %d\n", fake_nid);
131 return 1;
132 }
133 return 0;
134}
135
136static void __init reset_numa_cpu_lookup_table(void)
137{
138 unsigned int cpu;
139
140 for_each_possible_cpu(cpu)
141 numa_cpu_lookup_table[cpu] = -1;
142}
143
144void map_cpu_to_node(int cpu, int node)
145{
146 update_numa_cpu_lookup_table(cpu, node);
147
148 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) {
149 pr_debug("adding cpu %d to node %d\n", cpu, node);
150 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
151 }
152}
153
154#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
155void unmap_cpu_from_node(unsigned long cpu)
156{
157 int node = numa_cpu_lookup_table[cpu];
158
159 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
160 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
161 pr_debug("removing cpu %lu from node %d\n", cpu, node);
162 } else {
163 pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node);
164 }
165}
166#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
167
168static int __associativity_to_nid(const __be32 *associativity,
169 int max_array_sz)
170{
171 int nid;
172 /*
173 * primary_domain_index is 1 based array index.
174 */
175 int index = primary_domain_index - 1;
176
177 if (!numa_enabled || index >= max_array_sz)
178 return NUMA_NO_NODE;
179
180 nid = of_read_number(&associativity[index], 1);
181
182 /* POWER4 LPAR uses 0xffff as invalid node */
183 if (nid == 0xffff || nid >= nr_node_ids)
184 nid = NUMA_NO_NODE;
185 return nid;
186}
187/*
188 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
189 * info is found.
190 */
191static int associativity_to_nid(const __be32 *associativity)
192{
193 int array_sz = of_read_number(associativity, 1);
194
195 /* Skip the first element in the associativity array */
196 return __associativity_to_nid((associativity + 1), array_sz);
197}
198
199static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
200{
201 int dist;
202 int node1, node2;
203
204 node1 = associativity_to_nid(cpu1_assoc);
205 node2 = associativity_to_nid(cpu2_assoc);
206
207 dist = numa_distance_table[node1][node2];
208 if (dist <= LOCAL_DISTANCE)
209 return 0;
210 else if (dist <= REMOTE_DISTANCE)
211 return 1;
212 else
213 return 2;
214}
215
216static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
217{
218 int dist = 0;
219
220 int i, index;
221
222 for (i = 0; i < distance_ref_points_depth; i++) {
223 index = be32_to_cpu(distance_ref_points[i]);
224 if (cpu1_assoc[index] == cpu2_assoc[index])
225 break;
226 dist++;
227 }
228
229 return dist;
230}
231
232int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
233{
234 /* We should not get called with FORM0 */
235 VM_WARN_ON(affinity_form == FORM0_AFFINITY);
236 if (affinity_form == FORM1_AFFINITY)
237 return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
238 return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
239}
240
241/* must hold reference to node during call */
242static const __be32 *of_get_associativity(struct device_node *dev)
243{
244 return of_get_property(dev, "ibm,associativity", NULL);
245}
246
247int __node_distance(int a, int b)
248{
249 int i;
250 int distance = LOCAL_DISTANCE;
251
252 if (affinity_form == FORM2_AFFINITY)
253 return numa_distance_table[a][b];
254 else if (affinity_form == FORM0_AFFINITY)
255 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
256
257 for (i = 0; i < distance_ref_points_depth; i++) {
258 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
259 break;
260
261 /* Double the distance for each NUMA level */
262 distance *= 2;
263 }
264
265 return distance;
266}
267EXPORT_SYMBOL(__node_distance);
268
269/* Returns the nid associated with the given device tree node,
270 * or -1 if not found.
271 */
272static int of_node_to_nid_single(struct device_node *device)
273{
274 int nid = NUMA_NO_NODE;
275 const __be32 *tmp;
276
277 tmp = of_get_associativity(device);
278 if (tmp)
279 nid = associativity_to_nid(tmp);
280 return nid;
281}
282
283/* Walk the device tree upwards, looking for an associativity id */
284int of_node_to_nid(struct device_node *device)
285{
286 int nid = NUMA_NO_NODE;
287
288 of_node_get(device);
289 while (device) {
290 nid = of_node_to_nid_single(device);
291 if (nid != -1)
292 break;
293
294 device = of_get_next_parent(device);
295 }
296 of_node_put(device);
297
298 return nid;
299}
300EXPORT_SYMBOL(of_node_to_nid);
301
302static void __initialize_form1_numa_distance(const __be32 *associativity,
303 int max_array_sz)
304{
305 int i, nid;
306
307 if (affinity_form != FORM1_AFFINITY)
308 return;
309
310 nid = __associativity_to_nid(associativity, max_array_sz);
311 if (nid != NUMA_NO_NODE) {
312 for (i = 0; i < distance_ref_points_depth; i++) {
313 const __be32 *entry;
314 int index = be32_to_cpu(distance_ref_points[i]) - 1;
315
316 /*
317 * broken hierarchy, return with broken distance table
318 */
319 if (WARN(index >= max_array_sz, "Broken ibm,associativity property"))
320 return;
321
322 entry = &associativity[index];
323 distance_lookup_table[nid][i] = of_read_number(entry, 1);
324 }
325 }
326}
327
328static void initialize_form1_numa_distance(const __be32 *associativity)
329{
330 int array_sz;
331
332 array_sz = of_read_number(associativity, 1);
333 /* Skip the first element in the associativity array */
334 __initialize_form1_numa_distance(associativity + 1, array_sz);
335}
336
337/*
338 * Used to update distance information w.r.t newly added node.
339 */
340void update_numa_distance(struct device_node *node)
341{
342 int nid;
343
344 if (affinity_form == FORM0_AFFINITY)
345 return;
346 else if (affinity_form == FORM1_AFFINITY) {
347 const __be32 *associativity;
348
349 associativity = of_get_associativity(node);
350 if (!associativity)
351 return;
352
353 initialize_form1_numa_distance(associativity);
354 return;
355 }
356
357 /* FORM2 affinity */
358 nid = of_node_to_nid_single(node);
359 if (nid == NUMA_NO_NODE)
360 return;
361
362 /*
363 * With FORM2 we expect NUMA distance of all possible NUMA
364 * nodes to be provided during boot.
365 */
366 WARN(numa_distance_table[nid][nid] == -1,
367 "NUMA distance details for node %d not provided\n", nid);
368}
369
370/*
371 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
372 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
373 */
374static void __init initialize_form2_numa_distance_lookup_table(void)
375{
376 int i, j;
377 struct device_node *root;
378 const __u8 *form2_distances;
379 const __be32 *numa_lookup_index;
380 int form2_distances_length;
381 int max_numa_index, distance_index;
382
383 if (firmware_has_feature(FW_FEATURE_OPAL))
384 root = of_find_node_by_path("/ibm,opal");
385 else
386 root = of_find_node_by_path("/rtas");
387 if (!root)
388 root = of_find_node_by_path("/");
389
390 numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
391 max_numa_index = of_read_number(&numa_lookup_index[0], 1);
392
393 /* first element of the array is the size and is encode-int */
394 form2_distances = of_get_property(root, "ibm,numa-distance-table", NULL);
395 form2_distances_length = of_read_number((const __be32 *)&form2_distances[0], 1);
396 /* Skip the size which is encoded int */
397 form2_distances += sizeof(__be32);
398
399 pr_debug("form2_distances_len = %d, numa_dist_indexes_len = %d\n",
400 form2_distances_length, max_numa_index);
401
402 for (i = 0; i < max_numa_index; i++)
403 /* +1 skip the max_numa_index in the property */
404 numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
405
406
407 if (form2_distances_length != max_numa_index * max_numa_index) {
408 WARN(1, "Wrong NUMA distance information\n");
409 form2_distances = NULL; // don't use it
410 }
411 distance_index = 0;
412 for (i = 0; i < max_numa_index; i++) {
413 for (j = 0; j < max_numa_index; j++) {
414 int nodeA = numa_id_index_table[i];
415 int nodeB = numa_id_index_table[j];
416 int dist;
417
418 if (form2_distances)
419 dist = form2_distances[distance_index++];
420 else if (nodeA == nodeB)
421 dist = LOCAL_DISTANCE;
422 else
423 dist = REMOTE_DISTANCE;
424 numa_distance_table[nodeA][nodeB] = dist;
425 pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, dist);
426 }
427 }
428
429 of_node_put(root);
430}
431
432static int __init find_primary_domain_index(void)
433{
434 int index;
435 struct device_node *root;
436
437 /*
438 * Check for which form of affinity.
439 */
440 if (firmware_has_feature(FW_FEATURE_OPAL)) {
441 affinity_form = FORM1_AFFINITY;
442 } else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
443 pr_debug("Using form 2 affinity\n");
444 affinity_form = FORM2_AFFINITY;
445 } else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
446 pr_debug("Using form 1 affinity\n");
447 affinity_form = FORM1_AFFINITY;
448 } else
449 affinity_form = FORM0_AFFINITY;
450
451 if (firmware_has_feature(FW_FEATURE_OPAL))
452 root = of_find_node_by_path("/ibm,opal");
453 else
454 root = of_find_node_by_path("/rtas");
455 if (!root)
456 root = of_find_node_by_path("/");
457
458 /*
459 * This property is a set of 32-bit integers, each representing
460 * an index into the ibm,associativity nodes.
461 *
462 * With form 0 affinity the first integer is for an SMP configuration
463 * (should be all 0's) and the second is for a normal NUMA
464 * configuration. We have only one level of NUMA.
465 *
466 * With form 1 affinity the first integer is the most significant
467 * NUMA boundary and the following are progressively less significant
468 * boundaries. There can be more than one level of NUMA.
469 */
470 distance_ref_points = of_get_property(root,
471 "ibm,associativity-reference-points",
472 &distance_ref_points_depth);
473
474 if (!distance_ref_points) {
475 pr_debug("ibm,associativity-reference-points not found.\n");
476 goto err;
477 }
478
479 distance_ref_points_depth /= sizeof(int);
480 if (affinity_form == FORM0_AFFINITY) {
481 if (distance_ref_points_depth < 2) {
482 pr_warn("short ibm,associativity-reference-points\n");
483 goto err;
484 }
485
486 index = of_read_number(&distance_ref_points[1], 1);
487 } else {
488 /*
489 * Both FORM1 and FORM2 affinity find the primary domain details
490 * at the same offset.
491 */
492 index = of_read_number(distance_ref_points, 1);
493 }
494 /*
495 * Warn and cap if the hardware supports more than
496 * MAX_DISTANCE_REF_POINTS domains.
497 */
498 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
499 pr_warn("distance array capped at %d entries\n",
500 MAX_DISTANCE_REF_POINTS);
501 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
502 }
503
504 of_node_put(root);
505 return index;
506
507err:
508 of_node_put(root);
509 return -1;
510}
511
512static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
513{
514 struct device_node *memory = NULL;
515
516 memory = of_find_node_by_type(memory, "memory");
517 if (!memory)
518 panic("numa.c: No memory nodes found!");
519
520 *n_addr_cells = of_n_addr_cells(memory);
521 *n_size_cells = of_n_size_cells(memory);
522 of_node_put(memory);
523}
524
525static unsigned long read_n_cells(int n, const __be32 **buf)
526{
527 unsigned long result = 0;
528
529 while (n--) {
530 result = (result << 32) | of_read_number(*buf, 1);
531 (*buf)++;
532 }
533 return result;
534}
535
536struct assoc_arrays {
537 u32 n_arrays;
538 u32 array_sz;
539 const __be32 *arrays;
540};
541
542/*
543 * Retrieve and validate the list of associativity arrays for drconf
544 * memory from the ibm,associativity-lookup-arrays property of the
545 * device tree..
546 *
547 * The layout of the ibm,associativity-lookup-arrays property is a number N
548 * indicating the number of associativity arrays, followed by a number M
549 * indicating the size of each associativity array, followed by a list
550 * of N associativity arrays.
551 */
552static int of_get_assoc_arrays(struct assoc_arrays *aa)
553{
554 struct device_node *memory;
555 const __be32 *prop;
556 u32 len;
557
558 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
559 if (!memory)
560 return -1;
561
562 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
563 if (!prop || len < 2 * sizeof(unsigned int)) {
564 of_node_put(memory);
565 return -1;
566 }
567
568 aa->n_arrays = of_read_number(prop++, 1);
569 aa->array_sz = of_read_number(prop++, 1);
570
571 of_node_put(memory);
572
573 /* Now that we know the number of arrays and size of each array,
574 * revalidate the size of the property read in.
575 */
576 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
577 return -1;
578
579 aa->arrays = prop;
580 return 0;
581}
582
583static int __init get_nid_and_numa_distance(struct drmem_lmb *lmb)
584{
585 struct assoc_arrays aa = { .arrays = NULL };
586 int default_nid = NUMA_NO_NODE;
587 int nid = default_nid;
588 int rc, index;
589
590 if ((primary_domain_index < 0) || !numa_enabled)
591 return default_nid;
592
593 rc = of_get_assoc_arrays(&aa);
594 if (rc)
595 return default_nid;
596
597 if (primary_domain_index <= aa.array_sz &&
598 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
599 const __be32 *associativity;
600
601 index = lmb->aa_index * aa.array_sz;
602 associativity = &aa.arrays[index];
603 nid = __associativity_to_nid(associativity, aa.array_sz);
604 if (nid > 0 && affinity_form == FORM1_AFFINITY) {
605 /*
606 * lookup array associativity entries have
607 * no length of the array as the first element.
608 */
609 __initialize_form1_numa_distance(associativity, aa.array_sz);
610 }
611 }
612 return nid;
613}
614
615/*
616 * This is like of_node_to_nid_single() for memory represented in the
617 * ibm,dynamic-reconfiguration-memory node.
618 */
619int of_drconf_to_nid_single(struct drmem_lmb *lmb)
620{
621 struct assoc_arrays aa = { .arrays = NULL };
622 int default_nid = NUMA_NO_NODE;
623 int nid = default_nid;
624 int rc, index;
625
626 if ((primary_domain_index < 0) || !numa_enabled)
627 return default_nid;
628
629 rc = of_get_assoc_arrays(&aa);
630 if (rc)
631 return default_nid;
632
633 if (primary_domain_index <= aa.array_sz &&
634 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
635 const __be32 *associativity;
636
637 index = lmb->aa_index * aa.array_sz;
638 associativity = &aa.arrays[index];
639 nid = __associativity_to_nid(associativity, aa.array_sz);
640 }
641 return nid;
642}
643
644#ifdef CONFIG_PPC_SPLPAR
645
646static int __vphn_get_associativity(long lcpu, __be32 *associativity)
647{
648 long rc, hwid;
649
650 /*
651 * On a shared lpar, device tree will not have node associativity.
652 * At this time lppaca, or its __old_status field may not be
653 * updated. Hence kernel cannot detect if its on a shared lpar. So
654 * request an explicit associativity irrespective of whether the
655 * lpar is shared or dedicated. Use the device tree property as a
656 * fallback. cpu_to_phys_id is only valid between
657 * smp_setup_cpu_maps() and smp_setup_pacas().
658 */
659 if (firmware_has_feature(FW_FEATURE_VPHN)) {
660 if (cpu_to_phys_id)
661 hwid = cpu_to_phys_id[lcpu];
662 else
663 hwid = get_hard_smp_processor_id(lcpu);
664
665 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
666 if (rc == H_SUCCESS)
667 return 0;
668 }
669
670 return -1;
671}
672
673static int vphn_get_nid(long lcpu)
674{
675 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
676
677
678 if (!__vphn_get_associativity(lcpu, associativity))
679 return associativity_to_nid(associativity);
680
681 return NUMA_NO_NODE;
682
683}
684#else
685
686static int __vphn_get_associativity(long lcpu, __be32 *associativity)
687{
688 return -1;
689}
690
691static int vphn_get_nid(long unused)
692{
693 return NUMA_NO_NODE;
694}
695#endif /* CONFIG_PPC_SPLPAR */
696
697/*
698 * Figure out to which domain a cpu belongs and stick it there.
699 * Return the id of the domain used.
700 */
701static int numa_setup_cpu(unsigned long lcpu)
702{
703 struct device_node *cpu;
704 int fcpu = cpu_first_thread_sibling(lcpu);
705 int nid = NUMA_NO_NODE;
706
707 if (!cpu_present(lcpu)) {
708 set_cpu_numa_node(lcpu, first_online_node);
709 return first_online_node;
710 }
711
712 /*
713 * If a valid cpu-to-node mapping is already available, use it
714 * directly instead of querying the firmware, since it represents
715 * the most recent mapping notified to us by the platform (eg: VPHN).
716 * Since cpu_to_node binding remains the same for all threads in the
717 * core. If a valid cpu-to-node mapping is already available, for
718 * the first thread in the core, use it.
719 */
720 nid = numa_cpu_lookup_table[fcpu];
721 if (nid >= 0) {
722 map_cpu_to_node(lcpu, nid);
723 return nid;
724 }
725
726 nid = vphn_get_nid(lcpu);
727 if (nid != NUMA_NO_NODE)
728 goto out_present;
729
730 cpu = of_get_cpu_node(lcpu, NULL);
731
732 if (!cpu) {
733 WARN_ON(1);
734 if (cpu_present(lcpu))
735 goto out_present;
736 else
737 goto out;
738 }
739
740 nid = of_node_to_nid_single(cpu);
741 of_node_put(cpu);
742
743out_present:
744 if (nid < 0 || !node_possible(nid))
745 nid = first_online_node;
746
747 /*
748 * Update for the first thread of the core. All threads of a core
749 * have to be part of the same node. This not only avoids querying
750 * for every other thread in the core, but always avoids a case
751 * where virtual node associativity change causes subsequent threads
752 * of a core to be associated with different nid. However if first
753 * thread is already online, expect it to have a valid mapping.
754 */
755 if (fcpu != lcpu) {
756 WARN_ON(cpu_online(fcpu));
757 map_cpu_to_node(fcpu, nid);
758 }
759
760 map_cpu_to_node(lcpu, nid);
761out:
762 return nid;
763}
764
765static void verify_cpu_node_mapping(int cpu, int node)
766{
767 int base, sibling, i;
768
769 /* Verify that all the threads in the core belong to the same node */
770 base = cpu_first_thread_sibling(cpu);
771
772 for (i = 0; i < threads_per_core; i++) {
773 sibling = base + i;
774
775 if (sibling == cpu || cpu_is_offline(sibling))
776 continue;
777
778 if (cpu_to_node(sibling) != node) {
779 WARN(1, "CPU thread siblings %d and %d don't belong"
780 " to the same node!\n", cpu, sibling);
781 break;
782 }
783 }
784}
785
786/* Must run before sched domains notifier. */
787static int ppc_numa_cpu_prepare(unsigned int cpu)
788{
789 int nid;
790
791 nid = numa_setup_cpu(cpu);
792 verify_cpu_node_mapping(cpu, nid);
793 return 0;
794}
795
796static int ppc_numa_cpu_dead(unsigned int cpu)
797{
798 return 0;
799}
800
801/*
802 * Check and possibly modify a memory region to enforce the memory limit.
803 *
804 * Returns the size the region should have to enforce the memory limit.
805 * This will either be the original value of size, a truncated value,
806 * or zero. If the returned value of size is 0 the region should be
807 * discarded as it lies wholly above the memory limit.
808 */
809static unsigned long __init numa_enforce_memory_limit(unsigned long start,
810 unsigned long size)
811{
812 /*
813 * We use memblock_end_of_DRAM() in here instead of memory_limit because
814 * we've already adjusted it for the limit and it takes care of
815 * having memory holes below the limit. Also, in the case of
816 * iommu_is_off, memory_limit is not set but is implicitly enforced.
817 */
818
819 if (start + size <= memblock_end_of_DRAM())
820 return size;
821
822 if (start >= memblock_end_of_DRAM())
823 return 0;
824
825 return memblock_end_of_DRAM() - start;
826}
827
828/*
829 * Reads the counter for a given entry in
830 * linux,drconf-usable-memory property
831 */
832static inline int __init read_usm_ranges(const __be32 **usm)
833{
834 /*
835 * For each lmb in ibm,dynamic-memory a corresponding
836 * entry in linux,drconf-usable-memory property contains
837 * a counter followed by that many (base, size) duple.
838 * read the counter from linux,drconf-usable-memory
839 */
840 return read_n_cells(n_mem_size_cells, usm);
841}
842
843/*
844 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
845 * node. This assumes n_mem_{addr,size}_cells have been set.
846 */
847static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
848 const __be32 **usm,
849 void *data)
850{
851 unsigned int ranges, is_kexec_kdump = 0;
852 unsigned long base, size, sz;
853 int nid;
854
855 /*
856 * Skip this block if the reserved bit is set in flags (0x80)
857 * or if the block is not assigned to this partition (0x8)
858 */
859 if ((lmb->flags & DRCONF_MEM_RESERVED)
860 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
861 return 0;
862
863 if (*usm)
864 is_kexec_kdump = 1;
865
866 base = lmb->base_addr;
867 size = drmem_lmb_size();
868 ranges = 1;
869
870 if (is_kexec_kdump) {
871 ranges = read_usm_ranges(usm);
872 if (!ranges) /* there are no (base, size) duple */
873 return 0;
874 }
875
876 do {
877 if (is_kexec_kdump) {
878 base = read_n_cells(n_mem_addr_cells, usm);
879 size = read_n_cells(n_mem_size_cells, usm);
880 }
881
882 nid = get_nid_and_numa_distance(lmb);
883 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
884 &nid);
885 node_set_online(nid);
886 sz = numa_enforce_memory_limit(base, size);
887 if (sz)
888 memblock_set_node(base, sz, &memblock.memory, nid);
889 } while (--ranges);
890
891 return 0;
892}
893
894static int __init parse_numa_properties(void)
895{
896 struct device_node *memory;
897 int default_nid = 0;
898 unsigned long i;
899 const __be32 *associativity;
900
901 if (numa_enabled == 0) {
902 pr_warn("disabled by user\n");
903 return -1;
904 }
905
906 primary_domain_index = find_primary_domain_index();
907
908 if (primary_domain_index < 0) {
909 /*
910 * if we fail to parse primary_domain_index from device tree
911 * mark the numa disabled, boot with numa disabled.
912 */
913 numa_enabled = false;
914 return primary_domain_index;
915 }
916
917 pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index);
918
919 /*
920 * If it is FORM2 initialize the distance table here.
921 */
922 if (affinity_form == FORM2_AFFINITY)
923 initialize_form2_numa_distance_lookup_table();
924
925 /*
926 * Even though we connect cpus to numa domains later in SMP
927 * init, we need to know the node ids now. This is because
928 * each node to be onlined must have NODE_DATA etc backing it.
929 */
930 for_each_present_cpu(i) {
931 __be32 vphn_assoc[VPHN_ASSOC_BUFSIZE];
932 struct device_node *cpu;
933 int nid = NUMA_NO_NODE;
934
935 memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32));
936
937 if (__vphn_get_associativity(i, vphn_assoc) == 0) {
938 nid = associativity_to_nid(vphn_assoc);
939 initialize_form1_numa_distance(vphn_assoc);
940 } else {
941
942 /*
943 * Don't fall back to default_nid yet -- we will plug
944 * cpus into nodes once the memory scan has discovered
945 * the topology.
946 */
947 cpu = of_get_cpu_node(i, NULL);
948 BUG_ON(!cpu);
949
950 associativity = of_get_associativity(cpu);
951 if (associativity) {
952 nid = associativity_to_nid(associativity);
953 initialize_form1_numa_distance(associativity);
954 }
955 of_node_put(cpu);
956 }
957
958 /* node_set_online() is an UB if 'nid' is negative */
959 if (likely(nid >= 0))
960 node_set_online(nid);
961 }
962
963 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
964
965 for_each_node_by_type(memory, "memory") {
966 unsigned long start;
967 unsigned long size;
968 int nid;
969 int ranges;
970 const __be32 *memcell_buf;
971 unsigned int len;
972
973 memcell_buf = of_get_property(memory,
974 "linux,usable-memory", &len);
975 if (!memcell_buf || len <= 0)
976 memcell_buf = of_get_property(memory, "reg", &len);
977 if (!memcell_buf || len <= 0)
978 continue;
979
980 /* ranges in cell */
981 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
982new_range:
983 /* these are order-sensitive, and modify the buffer pointer */
984 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
985 size = read_n_cells(n_mem_size_cells, &memcell_buf);
986
987 /*
988 * Assumption: either all memory nodes or none will
989 * have associativity properties. If none, then
990 * everything goes to default_nid.
991 */
992 associativity = of_get_associativity(memory);
993 if (associativity) {
994 nid = associativity_to_nid(associativity);
995 initialize_form1_numa_distance(associativity);
996 } else
997 nid = default_nid;
998
999 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
1000 node_set_online(nid);
1001
1002 size = numa_enforce_memory_limit(start, size);
1003 if (size)
1004 memblock_set_node(start, size, &memblock.memory, nid);
1005
1006 if (--ranges)
1007 goto new_range;
1008 }
1009
1010 /*
1011 * Now do the same thing for each MEMBLOCK listed in the
1012 * ibm,dynamic-memory property in the
1013 * ibm,dynamic-reconfiguration-memory node.
1014 */
1015 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1016 if (memory) {
1017 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
1018 of_node_put(memory);
1019 }
1020
1021 return 0;
1022}
1023
1024static void __init setup_nonnuma(void)
1025{
1026 unsigned long top_of_ram = memblock_end_of_DRAM();
1027 unsigned long total_ram = memblock_phys_mem_size();
1028 unsigned long start_pfn, end_pfn;
1029 unsigned int nid = 0;
1030 int i;
1031
1032 pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram);
1033 pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20);
1034
1035 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1036 fake_numa_create_new_node(end_pfn, &nid);
1037 memblock_set_node(PFN_PHYS(start_pfn),
1038 PFN_PHYS(end_pfn - start_pfn),
1039 &memblock.memory, nid);
1040 node_set_online(nid);
1041 }
1042}
1043
1044void __init dump_numa_cpu_topology(void)
1045{
1046 unsigned int node;
1047 unsigned int cpu, count;
1048
1049 if (!numa_enabled)
1050 return;
1051
1052 for_each_online_node(node) {
1053 pr_info("Node %d CPUs:", node);
1054
1055 count = 0;
1056 /*
1057 * If we used a CPU iterator here we would miss printing
1058 * the holes in the cpumap.
1059 */
1060 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1061 if (cpumask_test_cpu(cpu,
1062 node_to_cpumask_map[node])) {
1063 if (count == 0)
1064 pr_cont(" %u", cpu);
1065 ++count;
1066 } else {
1067 if (count > 1)
1068 pr_cont("-%u", cpu - 1);
1069 count = 0;
1070 }
1071 }
1072
1073 if (count > 1)
1074 pr_cont("-%u", nr_cpu_ids - 1);
1075 pr_cont("\n");
1076 }
1077}
1078
1079/* Initialize NODE_DATA for a node on the local memory */
1080static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
1081{
1082 u64 spanned_pages = end_pfn - start_pfn;
1083 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
1084 u64 nd_pa;
1085 void *nd;
1086 int tnid;
1087
1088 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
1089 if (!nd_pa)
1090 panic("Cannot allocate %zu bytes for node %d data\n",
1091 nd_size, nid);
1092
1093 nd = __va(nd_pa);
1094
1095 /* report and initialize */
1096 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
1097 nd_pa, nd_pa + nd_size - 1);
1098 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
1099 if (tnid != nid)
1100 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
1101
1102 node_data[nid] = nd;
1103 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
1104 NODE_DATA(nid)->node_id = nid;
1105 NODE_DATA(nid)->node_start_pfn = start_pfn;
1106 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
1107}
1108
1109static void __init find_possible_nodes(void)
1110{
1111 struct device_node *rtas;
1112 const __be32 *domains = NULL;
1113 int prop_length, max_nodes;
1114 u32 i;
1115
1116 if (!numa_enabled)
1117 return;
1118
1119 rtas = of_find_node_by_path("/rtas");
1120 if (!rtas)
1121 return;
1122
1123 /*
1124 * ibm,current-associativity-domains is a fairly recent property. If
1125 * it doesn't exist, then fallback on ibm,max-associativity-domains.
1126 * Current denotes what the platform can support compared to max
1127 * which denotes what the Hypervisor can support.
1128 *
1129 * If the LPAR is migratable, new nodes might be activated after a LPM,
1130 * so we should consider the max number in that case.
1131 */
1132 if (!of_get_property(of_root, "ibm,migratable-partition", NULL))
1133 domains = of_get_property(rtas,
1134 "ibm,current-associativity-domains",
1135 &prop_length);
1136 if (!domains) {
1137 domains = of_get_property(rtas, "ibm,max-associativity-domains",
1138 &prop_length);
1139 if (!domains)
1140 goto out;
1141 }
1142
1143 max_nodes = of_read_number(&domains[primary_domain_index], 1);
1144 pr_info("Partition configured for %d NUMA nodes.\n", max_nodes);
1145
1146 for (i = 0; i < max_nodes; i++) {
1147 if (!node_possible(i))
1148 node_set(i, node_possible_map);
1149 }
1150
1151 prop_length /= sizeof(int);
1152 if (prop_length > primary_domain_index + 2)
1153 coregroup_enabled = 1;
1154
1155out:
1156 of_node_put(rtas);
1157}
1158
1159void __init mem_topology_setup(void)
1160{
1161 int cpu;
1162
1163 max_low_pfn = max_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1164 min_low_pfn = MEMORY_START >> PAGE_SHIFT;
1165
1166 /*
1167 * Linux/mm assumes node 0 to be online at boot. However this is not
1168 * true on PowerPC, where node 0 is similar to any other node, it
1169 * could be cpuless, memoryless node. So force node 0 to be offline
1170 * for now. This will prevent cpuless, memoryless node 0 showing up
1171 * unnecessarily as online. If a node has cpus or memory that need
1172 * to be online, then node will anyway be marked online.
1173 */
1174 node_set_offline(0);
1175
1176 if (parse_numa_properties())
1177 setup_nonnuma();
1178
1179 /*
1180 * Modify the set of possible NUMA nodes to reflect information
1181 * available about the set of online nodes, and the set of nodes
1182 * that we expect to make use of for this platform's affinity
1183 * calculations.
1184 */
1185 nodes_and(node_possible_map, node_possible_map, node_online_map);
1186
1187 find_possible_nodes();
1188
1189 setup_node_to_cpumask_map();
1190
1191 reset_numa_cpu_lookup_table();
1192
1193 for_each_possible_cpu(cpu) {
1194 /*
1195 * Powerpc with CONFIG_NUMA always used to have a node 0,
1196 * even if it was memoryless or cpuless. For all cpus that
1197 * are possible but not present, cpu_to_node() would point
1198 * to node 0. To remove a cpuless, memoryless dummy node,
1199 * powerpc need to make sure all possible but not present
1200 * cpu_to_node are set to a proper node.
1201 */
1202 numa_setup_cpu(cpu);
1203 }
1204}
1205
1206void __init initmem_init(void)
1207{
1208 int nid;
1209
1210 memblock_dump_all();
1211
1212 for_each_online_node(nid) {
1213 unsigned long start_pfn, end_pfn;
1214
1215 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1216 setup_node_data(nid, start_pfn, end_pfn);
1217 }
1218
1219 sparse_init();
1220
1221 /*
1222 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1223 * even before we online them, so that we can use cpu_to_{node,mem}
1224 * early in boot, cf. smp_prepare_cpus().
1225 * _nocalls() + manual invocation is used because cpuhp is not yet
1226 * initialized for the boot CPU.
1227 */
1228 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1229 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
1230}
1231
1232static int __init early_numa(char *p)
1233{
1234 if (!p)
1235 return 0;
1236
1237 if (strstr(p, "off"))
1238 numa_enabled = 0;
1239
1240 p = strstr(p, "fake=");
1241 if (p)
1242 cmdline = p + strlen("fake=");
1243
1244 return 0;
1245}
1246early_param("numa", early_numa);
1247
1248#ifdef CONFIG_MEMORY_HOTPLUG
1249/*
1250 * Find the node associated with a hot added memory section for
1251 * memory represented in the device tree by the property
1252 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1253 */
1254static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
1255{
1256 struct drmem_lmb *lmb;
1257 unsigned long lmb_size;
1258 int nid = NUMA_NO_NODE;
1259
1260 lmb_size = drmem_lmb_size();
1261
1262 for_each_drmem_lmb(lmb) {
1263 /* skip this block if it is reserved or not assigned to
1264 * this partition */
1265 if ((lmb->flags & DRCONF_MEM_RESERVED)
1266 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1267 continue;
1268
1269 if ((scn_addr < lmb->base_addr)
1270 || (scn_addr >= (lmb->base_addr + lmb_size)))
1271 continue;
1272
1273 nid = of_drconf_to_nid_single(lmb);
1274 break;
1275 }
1276
1277 return nid;
1278}
1279
1280/*
1281 * Find the node associated with a hot added memory section for memory
1282 * represented in the device tree as a node (i.e. memory@XXXX) for
1283 * each memblock.
1284 */
1285static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1286{
1287 struct device_node *memory;
1288 int nid = NUMA_NO_NODE;
1289
1290 for_each_node_by_type(memory, "memory") {
1291 unsigned long start, size;
1292 int ranges;
1293 const __be32 *memcell_buf;
1294 unsigned int len;
1295
1296 memcell_buf = of_get_property(memory, "reg", &len);
1297 if (!memcell_buf || len <= 0)
1298 continue;
1299
1300 /* ranges in cell */
1301 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1302
1303 while (ranges--) {
1304 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1305 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1306
1307 if ((scn_addr < start) || (scn_addr >= (start + size)))
1308 continue;
1309
1310 nid = of_node_to_nid_single(memory);
1311 break;
1312 }
1313
1314 if (nid >= 0)
1315 break;
1316 }
1317
1318 of_node_put(memory);
1319
1320 return nid;
1321}
1322
1323/*
1324 * Find the node associated with a hot added memory section. Section
1325 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1326 * sections are fully contained within a single MEMBLOCK.
1327 */
1328int hot_add_scn_to_nid(unsigned long scn_addr)
1329{
1330 struct device_node *memory = NULL;
1331 int nid;
1332
1333 if (!numa_enabled)
1334 return first_online_node;
1335
1336 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1337 if (memory) {
1338 nid = hot_add_drconf_scn_to_nid(scn_addr);
1339 of_node_put(memory);
1340 } else {
1341 nid = hot_add_node_scn_to_nid(scn_addr);
1342 }
1343
1344 if (nid < 0 || !node_possible(nid))
1345 nid = first_online_node;
1346
1347 return nid;
1348}
1349
1350static u64 hot_add_drconf_memory_max(void)
1351{
1352 struct device_node *memory = NULL;
1353 struct device_node *dn = NULL;
1354 const __be64 *lrdr = NULL;
1355
1356 dn = of_find_node_by_path("/rtas");
1357 if (dn) {
1358 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1359 of_node_put(dn);
1360 if (lrdr)
1361 return be64_to_cpup(lrdr);
1362 }
1363
1364 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1365 if (memory) {
1366 of_node_put(memory);
1367 return drmem_lmb_memory_max();
1368 }
1369 return 0;
1370}
1371
1372/*
1373 * memory_hotplug_max - return max address of memory that may be added
1374 *
1375 * This is currently only used on systems that support drconfig memory
1376 * hotplug.
1377 */
1378u64 memory_hotplug_max(void)
1379{
1380 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1381}
1382#endif /* CONFIG_MEMORY_HOTPLUG */
1383
1384/* Virtual Processor Home Node (VPHN) support */
1385#ifdef CONFIG_PPC_SPLPAR
1386static int topology_inited;
1387
1388/*
1389 * Retrieve the new associativity information for a virtual processor's
1390 * home node.
1391 */
1392static long vphn_get_associativity(unsigned long cpu,
1393 __be32 *associativity)
1394{
1395 long rc;
1396
1397 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1398 VPHN_FLAG_VCPU, associativity);
1399
1400 switch (rc) {
1401 case H_SUCCESS:
1402 pr_debug("VPHN hcall succeeded. Reset polling...\n");
1403 goto out;
1404
1405 case H_FUNCTION:
1406 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
1407 break;
1408 case H_HARDWARE:
1409 pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
1410 "preventing VPHN. Disabling polling...\n");
1411 break;
1412 case H_PARAMETER:
1413 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1414 "Disabling polling...\n");
1415 break;
1416 default:
1417 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1418 , rc);
1419 break;
1420 }
1421out:
1422 return rc;
1423}
1424
1425void find_and_update_cpu_nid(int cpu)
1426{
1427 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1428 int new_nid;
1429
1430 /* Use associativity from first thread for all siblings */
1431 if (vphn_get_associativity(cpu, associativity))
1432 return;
1433
1434 /* Do not have previous associativity, so find it now. */
1435 new_nid = associativity_to_nid(associativity);
1436
1437 if (new_nid < 0 || !node_possible(new_nid))
1438 new_nid = first_online_node;
1439 else
1440 // Associate node <-> cpu, so cpu_up() calls
1441 // try_online_node() on the right node.
1442 set_cpu_numa_node(cpu, new_nid);
1443
1444 pr_debug("%s:%d cpu %d nid %d\n", __func__, __LINE__, cpu, new_nid);
1445}
1446
1447int cpu_to_coregroup_id(int cpu)
1448{
1449 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1450 int index;
1451
1452 if (cpu < 0 || cpu > nr_cpu_ids)
1453 return -1;
1454
1455 if (!coregroup_enabled)
1456 goto out;
1457
1458 if (!firmware_has_feature(FW_FEATURE_VPHN))
1459 goto out;
1460
1461 if (vphn_get_associativity(cpu, associativity))
1462 goto out;
1463
1464 index = of_read_number(associativity, 1);
1465 if (index > primary_domain_index + 1)
1466 return of_read_number(&associativity[index - 1], 1);
1467
1468out:
1469 return cpu_to_core_id(cpu);
1470}
1471
1472static int topology_update_init(void)
1473{
1474 topology_inited = 1;
1475 return 0;
1476}
1477device_initcall(topology_update_init);
1478#endif /* CONFIG_PPC_SPLPAR */
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 <linux/memblock.h>
21#include <linux/of.h>
22#include <linux/pfn.h>
23#include <linux/cpuset.h>
24#include <linux/node.h>
25#include <asm/sparsemem.h>
26#include <asm/prom.h>
27#include <asm/system.h>
28#include <asm/smp.h>
29#include <asm/firmware.h>
30#include <asm/paca.h>
31#include <asm/hvcall.h>
32
33static int numa_enabled = 1;
34
35static char *cmdline __initdata;
36
37static int numa_debug;
38#define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
39
40int numa_cpu_lookup_table[NR_CPUS];
41cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
42struct pglist_data *node_data[MAX_NUMNODES];
43
44EXPORT_SYMBOL(numa_cpu_lookup_table);
45EXPORT_SYMBOL(node_to_cpumask_map);
46EXPORT_SYMBOL(node_data);
47
48static int min_common_depth;
49static int n_mem_addr_cells, n_mem_size_cells;
50static int form1_affinity;
51
52#define MAX_DISTANCE_REF_POINTS 4
53static int distance_ref_points_depth;
54static const unsigned int *distance_ref_points;
55static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
56
57/*
58 * Allocate node_to_cpumask_map based on number of available nodes
59 * Requires node_possible_map to be valid.
60 *
61 * Note: node_to_cpumask() is not valid until after this is done.
62 */
63static void __init setup_node_to_cpumask_map(void)
64{
65 unsigned int node, num = 0;
66
67 /* setup nr_node_ids if not done yet */
68 if (nr_node_ids == MAX_NUMNODES) {
69 for_each_node_mask(node, node_possible_map)
70 num = node;
71 nr_node_ids = num + 1;
72 }
73
74 /* allocate the map */
75 for (node = 0; node < nr_node_ids; node++)
76 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
77
78 /* cpumask_of_node() will now work */
79 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
80}
81
82static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
83 unsigned int *nid)
84{
85 unsigned long long mem;
86 char *p = cmdline;
87 static unsigned int fake_nid;
88 static unsigned long long curr_boundary;
89
90 /*
91 * Modify node id, iff we started creating NUMA nodes
92 * We want to continue from where we left of the last time
93 */
94 if (fake_nid)
95 *nid = fake_nid;
96 /*
97 * In case there are no more arguments to parse, the
98 * node_id should be the same as the last fake node id
99 * (we've handled this above).
100 */
101 if (!p)
102 return 0;
103
104 mem = memparse(p, &p);
105 if (!mem)
106 return 0;
107
108 if (mem < curr_boundary)
109 return 0;
110
111 curr_boundary = mem;
112
113 if ((end_pfn << PAGE_SHIFT) > mem) {
114 /*
115 * Skip commas and spaces
116 */
117 while (*p == ',' || *p == ' ' || *p == '\t')
118 p++;
119
120 cmdline = p;
121 fake_nid++;
122 *nid = fake_nid;
123 dbg("created new fake_node with id %d\n", fake_nid);
124 return 1;
125 }
126 return 0;
127}
128
129/*
130 * get_active_region_work_fn - A helper function for get_node_active_region
131 * Returns datax set to the start_pfn and end_pfn if they contain
132 * the initial value of datax->start_pfn between them
133 * @start_pfn: start page(inclusive) of region to check
134 * @end_pfn: end page(exclusive) of region to check
135 * @datax: comes in with ->start_pfn set to value to search for and
136 * goes out with active range if it contains it
137 * Returns 1 if search value is in range else 0
138 */
139static int __init get_active_region_work_fn(unsigned long start_pfn,
140 unsigned long end_pfn, void *datax)
141{
142 struct node_active_region *data;
143 data = (struct node_active_region *)datax;
144
145 if (start_pfn <= data->start_pfn && end_pfn > data->start_pfn) {
146 data->start_pfn = start_pfn;
147 data->end_pfn = end_pfn;
148 return 1;
149 }
150 return 0;
151
152}
153
154/*
155 * get_node_active_region - Return active region containing start_pfn
156 * Active range returned is empty if none found.
157 * @start_pfn: The page to return the region for.
158 * @node_ar: Returned set to the active region containing start_pfn
159 */
160static void __init get_node_active_region(unsigned long start_pfn,
161 struct node_active_region *node_ar)
162{
163 int nid = early_pfn_to_nid(start_pfn);
164
165 node_ar->nid = nid;
166 node_ar->start_pfn = start_pfn;
167 node_ar->end_pfn = start_pfn;
168 work_with_active_regions(nid, get_active_region_work_fn, node_ar);
169}
170
171static void map_cpu_to_node(int cpu, int node)
172{
173 numa_cpu_lookup_table[cpu] = node;
174
175 dbg("adding cpu %d to node %d\n", cpu, node);
176
177 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
178 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
179}
180
181#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
182static void unmap_cpu_from_node(unsigned long cpu)
183{
184 int node = numa_cpu_lookup_table[cpu];
185
186 dbg("removing cpu %lu from node %d\n", cpu, node);
187
188 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
189 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
190 } else {
191 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
192 cpu, node);
193 }
194}
195#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
196
197/* must hold reference to node during call */
198static const int *of_get_associativity(struct device_node *dev)
199{
200 return of_get_property(dev, "ibm,associativity", NULL);
201}
202
203/*
204 * Returns the property linux,drconf-usable-memory if
205 * it exists (the property exists only in kexec/kdump kernels,
206 * added by kexec-tools)
207 */
208static const u32 *of_get_usable_memory(struct device_node *memory)
209{
210 const u32 *prop;
211 u32 len;
212 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
213 if (!prop || len < sizeof(unsigned int))
214 return 0;
215 return prop;
216}
217
218int __node_distance(int a, int b)
219{
220 int i;
221 int distance = LOCAL_DISTANCE;
222
223 if (!form1_affinity)
224 return distance;
225
226 for (i = 0; i < distance_ref_points_depth; i++) {
227 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
228 break;
229
230 /* Double the distance for each NUMA level */
231 distance *= 2;
232 }
233
234 return distance;
235}
236
237static void initialize_distance_lookup_table(int nid,
238 const unsigned int *associativity)
239{
240 int i;
241
242 if (!form1_affinity)
243 return;
244
245 for (i = 0; i < distance_ref_points_depth; i++) {
246 distance_lookup_table[nid][i] =
247 associativity[distance_ref_points[i]];
248 }
249}
250
251/* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
252 * info is found.
253 */
254static int associativity_to_nid(const unsigned int *associativity)
255{
256 int nid = -1;
257
258 if (min_common_depth == -1)
259 goto out;
260
261 if (associativity[0] >= min_common_depth)
262 nid = associativity[min_common_depth];
263
264 /* POWER4 LPAR uses 0xffff as invalid node */
265 if (nid == 0xffff || nid >= MAX_NUMNODES)
266 nid = -1;
267
268 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
269 initialize_distance_lookup_table(nid, associativity);
270
271out:
272 return nid;
273}
274
275/* Returns the nid associated with the given device tree node,
276 * or -1 if not found.
277 */
278static int of_node_to_nid_single(struct device_node *device)
279{
280 int nid = -1;
281 const unsigned int *tmp;
282
283 tmp = of_get_associativity(device);
284 if (tmp)
285 nid = associativity_to_nid(tmp);
286 return nid;
287}
288
289/* Walk the device tree upwards, looking for an associativity id */
290int of_node_to_nid(struct device_node *device)
291{
292 struct device_node *tmp;
293 int nid = -1;
294
295 of_node_get(device);
296 while (device) {
297 nid = of_node_to_nid_single(device);
298 if (nid != -1)
299 break;
300
301 tmp = device;
302 device = of_get_parent(tmp);
303 of_node_put(tmp);
304 }
305 of_node_put(device);
306
307 return nid;
308}
309EXPORT_SYMBOL_GPL(of_node_to_nid);
310
311static int __init find_min_common_depth(void)
312{
313 int depth;
314 struct device_node *chosen;
315 struct device_node *root;
316 const char *vec5;
317
318 root = of_find_node_by_path("/rtas");
319 if (!root)
320 root = of_find_node_by_path("/");
321
322 /*
323 * This property is a set of 32-bit integers, each representing
324 * an index into the ibm,associativity nodes.
325 *
326 * With form 0 affinity the first integer is for an SMP configuration
327 * (should be all 0's) and the second is for a normal NUMA
328 * configuration. We have only one level of NUMA.
329 *
330 * With form 1 affinity the first integer is the most significant
331 * NUMA boundary and the following are progressively less significant
332 * boundaries. There can be more than one level of NUMA.
333 */
334 distance_ref_points = of_get_property(root,
335 "ibm,associativity-reference-points",
336 &distance_ref_points_depth);
337
338 if (!distance_ref_points) {
339 dbg("NUMA: ibm,associativity-reference-points not found.\n");
340 goto err;
341 }
342
343 distance_ref_points_depth /= sizeof(int);
344
345#define VEC5_AFFINITY_BYTE 5
346#define VEC5_AFFINITY 0x80
347 chosen = of_find_node_by_path("/chosen");
348 if (chosen) {
349 vec5 = of_get_property(chosen, "ibm,architecture-vec-5", NULL);
350 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] & VEC5_AFFINITY)) {
351 dbg("Using form 1 affinity\n");
352 form1_affinity = 1;
353 }
354 }
355
356 if (form1_affinity) {
357 depth = distance_ref_points[0];
358 } else {
359 if (distance_ref_points_depth < 2) {
360 printk(KERN_WARNING "NUMA: "
361 "short ibm,associativity-reference-points\n");
362 goto err;
363 }
364
365 depth = distance_ref_points[1];
366 }
367
368 /*
369 * Warn and cap if the hardware supports more than
370 * MAX_DISTANCE_REF_POINTS domains.
371 */
372 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
373 printk(KERN_WARNING "NUMA: distance array capped at "
374 "%d entries\n", MAX_DISTANCE_REF_POINTS);
375 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
376 }
377
378 of_node_put(root);
379 return depth;
380
381err:
382 of_node_put(root);
383 return -1;
384}
385
386static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
387{
388 struct device_node *memory = NULL;
389
390 memory = of_find_node_by_type(memory, "memory");
391 if (!memory)
392 panic("numa.c: No memory nodes found!");
393
394 *n_addr_cells = of_n_addr_cells(memory);
395 *n_size_cells = of_n_size_cells(memory);
396 of_node_put(memory);
397}
398
399static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
400{
401 unsigned long result = 0;
402
403 while (n--) {
404 result = (result << 32) | **buf;
405 (*buf)++;
406 }
407 return result;
408}
409
410struct of_drconf_cell {
411 u64 base_addr;
412 u32 drc_index;
413 u32 reserved;
414 u32 aa_index;
415 u32 flags;
416};
417
418#define DRCONF_MEM_ASSIGNED 0x00000008
419#define DRCONF_MEM_AI_INVALID 0x00000040
420#define DRCONF_MEM_RESERVED 0x00000080
421
422/*
423 * Read the next memblock list entry from the ibm,dynamic-memory property
424 * and return the information in the provided of_drconf_cell structure.
425 */
426static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
427{
428 const u32 *cp;
429
430 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
431
432 cp = *cellp;
433 drmem->drc_index = cp[0];
434 drmem->reserved = cp[1];
435 drmem->aa_index = cp[2];
436 drmem->flags = cp[3];
437
438 *cellp = cp + 4;
439}
440
441/*
442 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
443 *
444 * The layout of the ibm,dynamic-memory property is a number N of memblock
445 * list entries followed by N memblock list entries. Each memblock list entry
446 * contains information as laid out in the of_drconf_cell struct above.
447 */
448static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
449{
450 const u32 *prop;
451 u32 len, entries;
452
453 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
454 if (!prop || len < sizeof(unsigned int))
455 return 0;
456
457 entries = *prop++;
458
459 /* Now that we know the number of entries, revalidate the size
460 * of the property read in to ensure we have everything
461 */
462 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
463 return 0;
464
465 *dm = prop;
466 return entries;
467}
468
469/*
470 * Retrieve and validate the ibm,lmb-size property for drconf memory
471 * from the device tree.
472 */
473static u64 of_get_lmb_size(struct device_node *memory)
474{
475 const u32 *prop;
476 u32 len;
477
478 prop = of_get_property(memory, "ibm,lmb-size", &len);
479 if (!prop || len < sizeof(unsigned int))
480 return 0;
481
482 return read_n_cells(n_mem_size_cells, &prop);
483}
484
485struct assoc_arrays {
486 u32 n_arrays;
487 u32 array_sz;
488 const u32 *arrays;
489};
490
491/*
492 * Retrieve and validate the list of associativity arrays for drconf
493 * memory from the ibm,associativity-lookup-arrays property of the
494 * device tree..
495 *
496 * The layout of the ibm,associativity-lookup-arrays property is a number N
497 * indicating the number of associativity arrays, followed by a number M
498 * indicating the size of each associativity array, followed by a list
499 * of N associativity arrays.
500 */
501static int of_get_assoc_arrays(struct device_node *memory,
502 struct assoc_arrays *aa)
503{
504 const u32 *prop;
505 u32 len;
506
507 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
508 if (!prop || len < 2 * sizeof(unsigned int))
509 return -1;
510
511 aa->n_arrays = *prop++;
512 aa->array_sz = *prop++;
513
514 /* Now that we know the number of arrrays and size of each array,
515 * revalidate the size of the property read in.
516 */
517 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
518 return -1;
519
520 aa->arrays = prop;
521 return 0;
522}
523
524/*
525 * This is like of_node_to_nid_single() for memory represented in the
526 * ibm,dynamic-reconfiguration-memory node.
527 */
528static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
529 struct assoc_arrays *aa)
530{
531 int default_nid = 0;
532 int nid = default_nid;
533 int index;
534
535 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
536 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
537 drmem->aa_index < aa->n_arrays) {
538 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
539 nid = aa->arrays[index];
540
541 if (nid == 0xffff || nid >= MAX_NUMNODES)
542 nid = default_nid;
543 }
544
545 return nid;
546}
547
548/*
549 * Figure out to which domain a cpu belongs and stick it there.
550 * Return the id of the domain used.
551 */
552static int __cpuinit numa_setup_cpu(unsigned long lcpu)
553{
554 int nid = 0;
555 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
556
557 if (!cpu) {
558 WARN_ON(1);
559 goto out;
560 }
561
562 nid = of_node_to_nid_single(cpu);
563
564 if (nid < 0 || !node_online(nid))
565 nid = first_online_node;
566out:
567 map_cpu_to_node(lcpu, nid);
568
569 of_node_put(cpu);
570
571 return nid;
572}
573
574static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
575 unsigned long action,
576 void *hcpu)
577{
578 unsigned long lcpu = (unsigned long)hcpu;
579 int ret = NOTIFY_DONE;
580
581 switch (action) {
582 case CPU_UP_PREPARE:
583 case CPU_UP_PREPARE_FROZEN:
584 numa_setup_cpu(lcpu);
585 ret = NOTIFY_OK;
586 break;
587#ifdef CONFIG_HOTPLUG_CPU
588 case CPU_DEAD:
589 case CPU_DEAD_FROZEN:
590 case CPU_UP_CANCELED:
591 case CPU_UP_CANCELED_FROZEN:
592 unmap_cpu_from_node(lcpu);
593 break;
594 ret = NOTIFY_OK;
595#endif
596 }
597 return ret;
598}
599
600/*
601 * Check and possibly modify a memory region to enforce the memory limit.
602 *
603 * Returns the size the region should have to enforce the memory limit.
604 * This will either be the original value of size, a truncated value,
605 * or zero. If the returned value of size is 0 the region should be
606 * discarded as it lies wholly above the memory limit.
607 */
608static unsigned long __init numa_enforce_memory_limit(unsigned long start,
609 unsigned long size)
610{
611 /*
612 * We use memblock_end_of_DRAM() in here instead of memory_limit because
613 * we've already adjusted it for the limit and it takes care of
614 * having memory holes below the limit. Also, in the case of
615 * iommu_is_off, memory_limit is not set but is implicitly enforced.
616 */
617
618 if (start + size <= memblock_end_of_DRAM())
619 return size;
620
621 if (start >= memblock_end_of_DRAM())
622 return 0;
623
624 return memblock_end_of_DRAM() - start;
625}
626
627/*
628 * Reads the counter for a given entry in
629 * linux,drconf-usable-memory property
630 */
631static inline int __init read_usm_ranges(const u32 **usm)
632{
633 /*
634 * For each lmb in ibm,dynamic-memory a corresponding
635 * entry in linux,drconf-usable-memory property contains
636 * a counter followed by that many (base, size) duple.
637 * read the counter from linux,drconf-usable-memory
638 */
639 return read_n_cells(n_mem_size_cells, usm);
640}
641
642/*
643 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
644 * node. This assumes n_mem_{addr,size}_cells have been set.
645 */
646static void __init parse_drconf_memory(struct device_node *memory)
647{
648 const u32 *dm, *usm;
649 unsigned int n, rc, ranges, is_kexec_kdump = 0;
650 unsigned long lmb_size, base, size, sz;
651 int nid;
652 struct assoc_arrays aa;
653
654 n = of_get_drconf_memory(memory, &dm);
655 if (!n)
656 return;
657
658 lmb_size = of_get_lmb_size(memory);
659 if (!lmb_size)
660 return;
661
662 rc = of_get_assoc_arrays(memory, &aa);
663 if (rc)
664 return;
665
666 /* check if this is a kexec/kdump kernel */
667 usm = of_get_usable_memory(memory);
668 if (usm != NULL)
669 is_kexec_kdump = 1;
670
671 for (; n != 0; --n) {
672 struct of_drconf_cell drmem;
673
674 read_drconf_cell(&drmem, &dm);
675
676 /* skip this block if the reserved bit is set in flags (0x80)
677 or if the block is not assigned to this partition (0x8) */
678 if ((drmem.flags & DRCONF_MEM_RESERVED)
679 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
680 continue;
681
682 base = drmem.base_addr;
683 size = lmb_size;
684 ranges = 1;
685
686 if (is_kexec_kdump) {
687 ranges = read_usm_ranges(&usm);
688 if (!ranges) /* there are no (base, size) duple */
689 continue;
690 }
691 do {
692 if (is_kexec_kdump) {
693 base = read_n_cells(n_mem_addr_cells, &usm);
694 size = read_n_cells(n_mem_size_cells, &usm);
695 }
696 nid = of_drconf_to_nid_single(&drmem, &aa);
697 fake_numa_create_new_node(
698 ((base + size) >> PAGE_SHIFT),
699 &nid);
700 node_set_online(nid);
701 sz = numa_enforce_memory_limit(base, size);
702 if (sz)
703 add_active_range(nid, base >> PAGE_SHIFT,
704 (base >> PAGE_SHIFT)
705 + (sz >> PAGE_SHIFT));
706 } while (--ranges);
707 }
708}
709
710static int __init parse_numa_properties(void)
711{
712 struct device_node *cpu = NULL;
713 struct device_node *memory = NULL;
714 int default_nid = 0;
715 unsigned long i;
716
717 if (numa_enabled == 0) {
718 printk(KERN_WARNING "NUMA disabled by user\n");
719 return -1;
720 }
721
722 min_common_depth = find_min_common_depth();
723
724 if (min_common_depth < 0)
725 return min_common_depth;
726
727 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
728
729 /*
730 * Even though we connect cpus to numa domains later in SMP
731 * init, we need to know the node ids now. This is because
732 * each node to be onlined must have NODE_DATA etc backing it.
733 */
734 for_each_present_cpu(i) {
735 int nid;
736
737 cpu = of_get_cpu_node(i, NULL);
738 BUG_ON(!cpu);
739 nid = of_node_to_nid_single(cpu);
740 of_node_put(cpu);
741
742 /*
743 * Don't fall back to default_nid yet -- we will plug
744 * cpus into nodes once the memory scan has discovered
745 * the topology.
746 */
747 if (nid < 0)
748 continue;
749 node_set_online(nid);
750 }
751
752 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
753 memory = NULL;
754 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
755 unsigned long start;
756 unsigned long size;
757 int nid;
758 int ranges;
759 const unsigned int *memcell_buf;
760 unsigned int len;
761
762 memcell_buf = of_get_property(memory,
763 "linux,usable-memory", &len);
764 if (!memcell_buf || len <= 0)
765 memcell_buf = of_get_property(memory, "reg", &len);
766 if (!memcell_buf || len <= 0)
767 continue;
768
769 /* ranges in cell */
770 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
771new_range:
772 /* these are order-sensitive, and modify the buffer pointer */
773 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
774 size = read_n_cells(n_mem_size_cells, &memcell_buf);
775
776 /*
777 * Assumption: either all memory nodes or none will
778 * have associativity properties. If none, then
779 * everything goes to default_nid.
780 */
781 nid = of_node_to_nid_single(memory);
782 if (nid < 0)
783 nid = default_nid;
784
785 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
786 node_set_online(nid);
787
788 if (!(size = numa_enforce_memory_limit(start, size))) {
789 if (--ranges)
790 goto new_range;
791 else
792 continue;
793 }
794
795 add_active_range(nid, start >> PAGE_SHIFT,
796 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
797
798 if (--ranges)
799 goto new_range;
800 }
801
802 /*
803 * Now do the same thing for each MEMBLOCK listed in the ibm,dynamic-memory
804 * property in the ibm,dynamic-reconfiguration-memory node.
805 */
806 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
807 if (memory)
808 parse_drconf_memory(memory);
809
810 return 0;
811}
812
813static void __init setup_nonnuma(void)
814{
815 unsigned long top_of_ram = memblock_end_of_DRAM();
816 unsigned long total_ram = memblock_phys_mem_size();
817 unsigned long start_pfn, end_pfn;
818 unsigned int nid = 0;
819 struct memblock_region *reg;
820
821 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
822 top_of_ram, total_ram);
823 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
824 (top_of_ram - total_ram) >> 20);
825
826 for_each_memblock(memory, reg) {
827 start_pfn = memblock_region_memory_base_pfn(reg);
828 end_pfn = memblock_region_memory_end_pfn(reg);
829
830 fake_numa_create_new_node(end_pfn, &nid);
831 add_active_range(nid, start_pfn, end_pfn);
832 node_set_online(nid);
833 }
834}
835
836void __init dump_numa_cpu_topology(void)
837{
838 unsigned int node;
839 unsigned int cpu, count;
840
841 if (min_common_depth == -1 || !numa_enabled)
842 return;
843
844 for_each_online_node(node) {
845 printk(KERN_DEBUG "Node %d CPUs:", node);
846
847 count = 0;
848 /*
849 * If we used a CPU iterator here we would miss printing
850 * the holes in the cpumap.
851 */
852 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
853 if (cpumask_test_cpu(cpu,
854 node_to_cpumask_map[node])) {
855 if (count == 0)
856 printk(" %u", cpu);
857 ++count;
858 } else {
859 if (count > 1)
860 printk("-%u", cpu - 1);
861 count = 0;
862 }
863 }
864
865 if (count > 1)
866 printk("-%u", nr_cpu_ids - 1);
867 printk("\n");
868 }
869}
870
871static void __init dump_numa_memory_topology(void)
872{
873 unsigned int node;
874 unsigned int count;
875
876 if (min_common_depth == -1 || !numa_enabled)
877 return;
878
879 for_each_online_node(node) {
880 unsigned long i;
881
882 printk(KERN_DEBUG "Node %d Memory:", node);
883
884 count = 0;
885
886 for (i = 0; i < memblock_end_of_DRAM();
887 i += (1 << SECTION_SIZE_BITS)) {
888 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
889 if (count == 0)
890 printk(" 0x%lx", i);
891 ++count;
892 } else {
893 if (count > 0)
894 printk("-0x%lx", i);
895 count = 0;
896 }
897 }
898
899 if (count > 0)
900 printk("-0x%lx", i);
901 printk("\n");
902 }
903}
904
905/*
906 * Allocate some memory, satisfying the memblock or bootmem allocator where
907 * required. nid is the preferred node and end is the physical address of
908 * the highest address in the node.
909 *
910 * Returns the virtual address of the memory.
911 */
912static void __init *careful_zallocation(int nid, unsigned long size,
913 unsigned long align,
914 unsigned long end_pfn)
915{
916 void *ret;
917 int new_nid;
918 unsigned long ret_paddr;
919
920 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
921
922 /* retry over all memory */
923 if (!ret_paddr)
924 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
925
926 if (!ret_paddr)
927 panic("numa.c: cannot allocate %lu bytes for node %d",
928 size, nid);
929
930 ret = __va(ret_paddr);
931
932 /*
933 * We initialize the nodes in numeric order: 0, 1, 2...
934 * and hand over control from the MEMBLOCK allocator to the
935 * bootmem allocator. If this function is called for
936 * node 5, then we know that all nodes <5 are using the
937 * bootmem allocator instead of the MEMBLOCK allocator.
938 *
939 * So, check the nid from which this allocation came
940 * and double check to see if we need to use bootmem
941 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
942 * since it would be useless.
943 */
944 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
945 if (new_nid < nid) {
946 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
947 size, align, 0);
948
949 dbg("alloc_bootmem %p %lx\n", ret, size);
950 }
951
952 memset(ret, 0, size);
953 return ret;
954}
955
956static struct notifier_block __cpuinitdata ppc64_numa_nb = {
957 .notifier_call = cpu_numa_callback,
958 .priority = 1 /* Must run before sched domains notifier. */
959};
960
961static void mark_reserved_regions_for_nid(int nid)
962{
963 struct pglist_data *node = NODE_DATA(nid);
964 struct memblock_region *reg;
965
966 for_each_memblock(reserved, reg) {
967 unsigned long physbase = reg->base;
968 unsigned long size = reg->size;
969 unsigned long start_pfn = physbase >> PAGE_SHIFT;
970 unsigned long end_pfn = PFN_UP(physbase + size);
971 struct node_active_region node_ar;
972 unsigned long node_end_pfn = node->node_start_pfn +
973 node->node_spanned_pages;
974
975 /*
976 * Check to make sure that this memblock.reserved area is
977 * within the bounds of the node that we care about.
978 * Checking the nid of the start and end points is not
979 * sufficient because the reserved area could span the
980 * entire node.
981 */
982 if (end_pfn <= node->node_start_pfn ||
983 start_pfn >= node_end_pfn)
984 continue;
985
986 get_node_active_region(start_pfn, &node_ar);
987 while (start_pfn < end_pfn &&
988 node_ar.start_pfn < node_ar.end_pfn) {
989 unsigned long reserve_size = size;
990 /*
991 * if reserved region extends past active region
992 * then trim size to active region
993 */
994 if (end_pfn > node_ar.end_pfn)
995 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
996 - physbase;
997 /*
998 * Only worry about *this* node, others may not
999 * yet have valid NODE_DATA().
1000 */
1001 if (node_ar.nid == nid) {
1002 dbg("reserve_bootmem %lx %lx nid=%d\n",
1003 physbase, reserve_size, node_ar.nid);
1004 reserve_bootmem_node(NODE_DATA(node_ar.nid),
1005 physbase, reserve_size,
1006 BOOTMEM_DEFAULT);
1007 }
1008 /*
1009 * if reserved region is contained in the active region
1010 * then done.
1011 */
1012 if (end_pfn <= node_ar.end_pfn)
1013 break;
1014
1015 /*
1016 * reserved region extends past the active region
1017 * get next active region that contains this
1018 * reserved region
1019 */
1020 start_pfn = node_ar.end_pfn;
1021 physbase = start_pfn << PAGE_SHIFT;
1022 size = size - reserve_size;
1023 get_node_active_region(start_pfn, &node_ar);
1024 }
1025 }
1026}
1027
1028
1029void __init do_init_bootmem(void)
1030{
1031 int nid;
1032
1033 min_low_pfn = 0;
1034 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1035 max_pfn = max_low_pfn;
1036
1037 if (parse_numa_properties())
1038 setup_nonnuma();
1039 else
1040 dump_numa_memory_topology();
1041
1042 for_each_online_node(nid) {
1043 unsigned long start_pfn, end_pfn;
1044 void *bootmem_vaddr;
1045 unsigned long bootmap_pages;
1046
1047 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1048
1049 /*
1050 * Allocate the node structure node local if possible
1051 *
1052 * Be careful moving this around, as it relies on all
1053 * previous nodes' bootmem to be initialized and have
1054 * all reserved areas marked.
1055 */
1056 NODE_DATA(nid) = careful_zallocation(nid,
1057 sizeof(struct pglist_data),
1058 SMP_CACHE_BYTES, end_pfn);
1059
1060 dbg("node %d\n", nid);
1061 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1062
1063 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1064 NODE_DATA(nid)->node_start_pfn = start_pfn;
1065 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1066
1067 if (NODE_DATA(nid)->node_spanned_pages == 0)
1068 continue;
1069
1070 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1071 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1072
1073 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1074 bootmem_vaddr = careful_zallocation(nid,
1075 bootmap_pages << PAGE_SHIFT,
1076 PAGE_SIZE, end_pfn);
1077
1078 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1079
1080 init_bootmem_node(NODE_DATA(nid),
1081 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1082 start_pfn, end_pfn);
1083
1084 free_bootmem_with_active_regions(nid, end_pfn);
1085 /*
1086 * Be very careful about moving this around. Future
1087 * calls to careful_zallocation() depend on this getting
1088 * done correctly.
1089 */
1090 mark_reserved_regions_for_nid(nid);
1091 sparse_memory_present_with_active_regions(nid);
1092 }
1093
1094 init_bootmem_done = 1;
1095
1096 /*
1097 * Now bootmem is initialised we can create the node to cpumask
1098 * lookup tables and setup the cpu callback to populate them.
1099 */
1100 setup_node_to_cpumask_map();
1101
1102 register_cpu_notifier(&ppc64_numa_nb);
1103 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1104 (void *)(unsigned long)boot_cpuid);
1105}
1106
1107void __init paging_init(void)
1108{
1109 unsigned long max_zone_pfns[MAX_NR_ZONES];
1110 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1111 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1112 free_area_init_nodes(max_zone_pfns);
1113}
1114
1115static int __init early_numa(char *p)
1116{
1117 if (!p)
1118 return 0;
1119
1120 if (strstr(p, "off"))
1121 numa_enabled = 0;
1122
1123 if (strstr(p, "debug"))
1124 numa_debug = 1;
1125
1126 p = strstr(p, "fake=");
1127 if (p)
1128 cmdline = p + strlen("fake=");
1129
1130 return 0;
1131}
1132early_param("numa", early_numa);
1133
1134#ifdef CONFIG_MEMORY_HOTPLUG
1135/*
1136 * Find the node associated with a hot added memory section for
1137 * memory represented in the device tree by the property
1138 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1139 */
1140static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1141 unsigned long scn_addr)
1142{
1143 const u32 *dm;
1144 unsigned int drconf_cell_cnt, rc;
1145 unsigned long lmb_size;
1146 struct assoc_arrays aa;
1147 int nid = -1;
1148
1149 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1150 if (!drconf_cell_cnt)
1151 return -1;
1152
1153 lmb_size = of_get_lmb_size(memory);
1154 if (!lmb_size)
1155 return -1;
1156
1157 rc = of_get_assoc_arrays(memory, &aa);
1158 if (rc)
1159 return -1;
1160
1161 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1162 struct of_drconf_cell drmem;
1163
1164 read_drconf_cell(&drmem, &dm);
1165
1166 /* skip this block if it is reserved or not assigned to
1167 * this partition */
1168 if ((drmem.flags & DRCONF_MEM_RESERVED)
1169 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1170 continue;
1171
1172 if ((scn_addr < drmem.base_addr)
1173 || (scn_addr >= (drmem.base_addr + lmb_size)))
1174 continue;
1175
1176 nid = of_drconf_to_nid_single(&drmem, &aa);
1177 break;
1178 }
1179
1180 return nid;
1181}
1182
1183/*
1184 * Find the node associated with a hot added memory section for memory
1185 * represented in the device tree as a node (i.e. memory@XXXX) for
1186 * each memblock.
1187 */
1188int hot_add_node_scn_to_nid(unsigned long scn_addr)
1189{
1190 struct device_node *memory = NULL;
1191 int nid = -1;
1192
1193 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
1194 unsigned long start, size;
1195 int ranges;
1196 const unsigned int *memcell_buf;
1197 unsigned int len;
1198
1199 memcell_buf = of_get_property(memory, "reg", &len);
1200 if (!memcell_buf || len <= 0)
1201 continue;
1202
1203 /* ranges in cell */
1204 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1205
1206 while (ranges--) {
1207 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1208 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1209
1210 if ((scn_addr < start) || (scn_addr >= (start + size)))
1211 continue;
1212
1213 nid = of_node_to_nid_single(memory);
1214 break;
1215 }
1216
1217 of_node_put(memory);
1218 if (nid >= 0)
1219 break;
1220 }
1221
1222 return nid;
1223}
1224
1225/*
1226 * Find the node associated with a hot added memory section. Section
1227 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1228 * sections are fully contained within a single MEMBLOCK.
1229 */
1230int hot_add_scn_to_nid(unsigned long scn_addr)
1231{
1232 struct device_node *memory = NULL;
1233 int nid, found = 0;
1234
1235 if (!numa_enabled || (min_common_depth < 0))
1236 return first_online_node;
1237
1238 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1239 if (memory) {
1240 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1241 of_node_put(memory);
1242 } else {
1243 nid = hot_add_node_scn_to_nid(scn_addr);
1244 }
1245
1246 if (nid < 0 || !node_online(nid))
1247 nid = first_online_node;
1248
1249 if (NODE_DATA(nid)->node_spanned_pages)
1250 return nid;
1251
1252 for_each_online_node(nid) {
1253 if (NODE_DATA(nid)->node_spanned_pages) {
1254 found = 1;
1255 break;
1256 }
1257 }
1258
1259 BUG_ON(!found);
1260 return nid;
1261}
1262
1263static u64 hot_add_drconf_memory_max(void)
1264{
1265 struct device_node *memory = NULL;
1266 unsigned int drconf_cell_cnt = 0;
1267 u64 lmb_size = 0;
1268 const u32 *dm = 0;
1269
1270 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1271 if (memory) {
1272 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1273 lmb_size = of_get_lmb_size(memory);
1274 of_node_put(memory);
1275 }
1276 return lmb_size * drconf_cell_cnt;
1277}
1278
1279/*
1280 * memory_hotplug_max - return max address of memory that may be added
1281 *
1282 * This is currently only used on systems that support drconfig memory
1283 * hotplug.
1284 */
1285u64 memory_hotplug_max(void)
1286{
1287 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1288}
1289#endif /* CONFIG_MEMORY_HOTPLUG */
1290
1291/* Virtual Processor Home Node (VPHN) support */
1292#ifdef CONFIG_PPC_SPLPAR
1293static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1294static cpumask_t cpu_associativity_changes_mask;
1295static int vphn_enabled;
1296static void set_topology_timer(void);
1297
1298/*
1299 * Store the current values of the associativity change counters in the
1300 * hypervisor.
1301 */
1302static void setup_cpu_associativity_change_counters(void)
1303{
1304 int cpu;
1305
1306 /* The VPHN feature supports a maximum of 8 reference points */
1307 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1308
1309 for_each_possible_cpu(cpu) {
1310 int i;
1311 u8 *counts = vphn_cpu_change_counts[cpu];
1312 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1313
1314 for (i = 0; i < distance_ref_points_depth; i++)
1315 counts[i] = hypervisor_counts[i];
1316 }
1317}
1318
1319/*
1320 * The hypervisor maintains a set of 8 associativity change counters in
1321 * the VPA of each cpu that correspond to the associativity levels in the
1322 * ibm,associativity-reference-points property. When an associativity
1323 * level changes, the corresponding counter is incremented.
1324 *
1325 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1326 * node associativity levels have changed.
1327 *
1328 * Returns the number of cpus with unhandled associativity changes.
1329 */
1330static int update_cpu_associativity_changes_mask(void)
1331{
1332 int cpu, nr_cpus = 0;
1333 cpumask_t *changes = &cpu_associativity_changes_mask;
1334
1335 cpumask_clear(changes);
1336
1337 for_each_possible_cpu(cpu) {
1338 int i, changed = 0;
1339 u8 *counts = vphn_cpu_change_counts[cpu];
1340 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1341
1342 for (i = 0; i < distance_ref_points_depth; i++) {
1343 if (hypervisor_counts[i] != counts[i]) {
1344 counts[i] = hypervisor_counts[i];
1345 changed = 1;
1346 }
1347 }
1348 if (changed) {
1349 cpumask_set_cpu(cpu, changes);
1350 nr_cpus++;
1351 }
1352 }
1353
1354 return nr_cpus;
1355}
1356
1357/*
1358 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1359 * the complete property we have to add the length in the first cell.
1360 */
1361#define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1362
1363/*
1364 * Convert the associativity domain numbers returned from the hypervisor
1365 * to the sequence they would appear in the ibm,associativity property.
1366 */
1367static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1368{
1369 int i, nr_assoc_doms = 0;
1370 const u16 *field = (const u16*) packed;
1371
1372#define VPHN_FIELD_UNUSED (0xffff)
1373#define VPHN_FIELD_MSB (0x8000)
1374#define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1375
1376 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1377 if (*field == VPHN_FIELD_UNUSED) {
1378 /* All significant fields processed, and remaining
1379 * fields contain the reserved value of all 1's.
1380 * Just store them.
1381 */
1382 unpacked[i] = *((u32*)field);
1383 field += 2;
1384 } else if (*field & VPHN_FIELD_MSB) {
1385 /* Data is in the lower 15 bits of this field */
1386 unpacked[i] = *field & VPHN_FIELD_MASK;
1387 field++;
1388 nr_assoc_doms++;
1389 } else {
1390 /* Data is in the lower 15 bits of this field
1391 * concatenated with the next 16 bit field
1392 */
1393 unpacked[i] = *((u32*)field);
1394 field += 2;
1395 nr_assoc_doms++;
1396 }
1397 }
1398
1399 /* The first cell contains the length of the property */
1400 unpacked[0] = nr_assoc_doms;
1401
1402 return nr_assoc_doms;
1403}
1404
1405/*
1406 * Retrieve the new associativity information for a virtual processor's
1407 * home node.
1408 */
1409static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1410{
1411 long rc;
1412 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1413 u64 flags = 1;
1414 int hwcpu = get_hard_smp_processor_id(cpu);
1415
1416 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1417 vphn_unpack_associativity(retbuf, associativity);
1418
1419 return rc;
1420}
1421
1422static long vphn_get_associativity(unsigned long cpu,
1423 unsigned int *associativity)
1424{
1425 long rc;
1426
1427 rc = hcall_vphn(cpu, associativity);
1428
1429 switch (rc) {
1430 case H_FUNCTION:
1431 printk(KERN_INFO
1432 "VPHN is not supported. Disabling polling...\n");
1433 stop_topology_update();
1434 break;
1435 case H_HARDWARE:
1436 printk(KERN_ERR
1437 "hcall_vphn() experienced a hardware fault "
1438 "preventing VPHN. Disabling polling...\n");
1439 stop_topology_update();
1440 }
1441
1442 return rc;
1443}
1444
1445/*
1446 * Update the node maps and sysfs entries for each cpu whose home node
1447 * has changed.
1448 */
1449int arch_update_cpu_topology(void)
1450{
1451 int cpu, nid, old_nid;
1452 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1453 struct sys_device *sysdev;
1454
1455 for_each_cpu(cpu,&cpu_associativity_changes_mask) {
1456 vphn_get_associativity(cpu, associativity);
1457 nid = associativity_to_nid(associativity);
1458
1459 if (nid < 0 || !node_online(nid))
1460 nid = first_online_node;
1461
1462 old_nid = numa_cpu_lookup_table[cpu];
1463
1464 /* Disable hotplug while we update the cpu
1465 * masks and sysfs.
1466 */
1467 get_online_cpus();
1468 unregister_cpu_under_node(cpu, old_nid);
1469 unmap_cpu_from_node(cpu);
1470 map_cpu_to_node(cpu, nid);
1471 register_cpu_under_node(cpu, nid);
1472 put_online_cpus();
1473
1474 sysdev = get_cpu_sysdev(cpu);
1475 if (sysdev)
1476 kobject_uevent(&sysdev->kobj, KOBJ_CHANGE);
1477 }
1478
1479 return 1;
1480}
1481
1482static void topology_work_fn(struct work_struct *work)
1483{
1484 rebuild_sched_domains();
1485}
1486static DECLARE_WORK(topology_work, topology_work_fn);
1487
1488void topology_schedule_update(void)
1489{
1490 schedule_work(&topology_work);
1491}
1492
1493static void topology_timer_fn(unsigned long ignored)
1494{
1495 if (!vphn_enabled)
1496 return;
1497 if (update_cpu_associativity_changes_mask() > 0)
1498 topology_schedule_update();
1499 set_topology_timer();
1500}
1501static struct timer_list topology_timer =
1502 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1503
1504static void set_topology_timer(void)
1505{
1506 topology_timer.data = 0;
1507 topology_timer.expires = jiffies + 60 * HZ;
1508 add_timer(&topology_timer);
1509}
1510
1511/*
1512 * Start polling for VPHN associativity changes.
1513 */
1514int start_topology_update(void)
1515{
1516 int rc = 0;
1517
1518 /* Disabled until races with load balancing are fixed */
1519 if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1520 get_lppaca()->shared_proc) {
1521 vphn_enabled = 1;
1522 setup_cpu_associativity_change_counters();
1523 init_timer_deferrable(&topology_timer);
1524 set_topology_timer();
1525 rc = 1;
1526 }
1527
1528 return rc;
1529}
1530__initcall(start_topology_update);
1531
1532/*
1533 * Disable polling for VPHN associativity changes.
1534 */
1535int stop_topology_update(void)
1536{
1537 vphn_enabled = 0;
1538 return del_timer_sync(&topology_timer);
1539}
1540#endif /* CONFIG_PPC_SPLPAR */