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/of_address.h>
20#include <linux/pfn.h>
21#include <linux/cpuset.h>
22#include <linux/node.h>
23#include <linux/stop_machine.h>
24#include <linux/proc_fs.h>
25#include <linux/seq_file.h>
26#include <linux/uaccess.h>
27#include <linux/slab.h>
28#include <asm/cputhreads.h>
29#include <asm/sparsemem.h>
30#include <asm/smp.h>
31#include <asm/topology.h>
32#include <asm/firmware.h>
33#include <asm/paca.h>
34#include <asm/hvcall.h>
35#include <asm/setup.h>
36#include <asm/vdso.h>
37#include <asm/vphn.h>
38#include <asm/drmem.h>
39
40static int numa_enabled = 1;
41
42static char *cmdline __initdata;
43
44int numa_cpu_lookup_table[NR_CPUS];
45cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
46struct pglist_data *node_data[MAX_NUMNODES];
47
48EXPORT_SYMBOL(numa_cpu_lookup_table);
49EXPORT_SYMBOL(node_to_cpumask_map);
50EXPORT_SYMBOL(node_data);
51
52static int primary_domain_index;
53static int n_mem_addr_cells, n_mem_size_cells;
54
55#define FORM0_AFFINITY 0
56#define FORM1_AFFINITY 1
57#define FORM2_AFFINITY 2
58static int affinity_form;
59
60#define MAX_DISTANCE_REF_POINTS 4
61static int distance_ref_points_depth;
62static const __be32 *distance_ref_points;
63static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
64static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
65 [0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
66};
67static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
68
69/*
70 * Allocate node_to_cpumask_map based on number of available nodes
71 * Requires node_possible_map to be valid.
72 *
73 * Note: cpumask_of_node() is not valid until after this is done.
74 */
75static void __init setup_node_to_cpumask_map(void)
76{
77 unsigned int node;
78
79 /* setup nr_node_ids if not done yet */
80 if (nr_node_ids == MAX_NUMNODES)
81 setup_nr_node_ids();
82
83 /* allocate the map */
84 for_each_node(node)
85 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
86
87 /* cpumask_of_node() will now work */
88 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
89}
90
91static int __init fake_numa_create_new_node(unsigned long end_pfn,
92 unsigned int *nid)
93{
94 unsigned long long mem;
95 char *p = cmdline;
96 static unsigned int fake_nid;
97 static unsigned long long curr_boundary;
98
99 /*
100 * Modify node id, iff we started creating NUMA nodes
101 * We want to continue from where we left of the last time
102 */
103 if (fake_nid)
104 *nid = fake_nid;
105 /*
106 * In case there are no more arguments to parse, the
107 * node_id should be the same as the last fake node id
108 * (we've handled this above).
109 */
110 if (!p)
111 return 0;
112
113 mem = memparse(p, &p);
114 if (!mem)
115 return 0;
116
117 if (mem < curr_boundary)
118 return 0;
119
120 curr_boundary = mem;
121
122 if ((end_pfn << PAGE_SHIFT) > mem) {
123 /*
124 * Skip commas and spaces
125 */
126 while (*p == ',' || *p == ' ' || *p == '\t')
127 p++;
128
129 cmdline = p;
130 fake_nid++;
131 *nid = fake_nid;
132 pr_debug("created new fake_node with id %d\n", fake_nid);
133 return 1;
134 }
135 return 0;
136}
137
138static void __init reset_numa_cpu_lookup_table(void)
139{
140 unsigned int cpu;
141
142 for_each_possible_cpu(cpu)
143 numa_cpu_lookup_table[cpu] = -1;
144}
145
146void map_cpu_to_node(int cpu, int node)
147{
148 update_numa_cpu_lookup_table(cpu, node);
149
150 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) {
151 pr_debug("adding cpu %d to node %d\n", cpu, node);
152 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
153 }
154}
155
156#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
157void unmap_cpu_from_node(unsigned long cpu)
158{
159 int node = numa_cpu_lookup_table[cpu];
160
161 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
162 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
163 pr_debug("removing cpu %lu from node %d\n", cpu, node);
164 } else {
165 pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node);
166 }
167}
168#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
169
170static int __associativity_to_nid(const __be32 *associativity,
171 int max_array_sz)
172{
173 int nid;
174 /*
175 * primary_domain_index is 1 based array index.
176 */
177 int index = primary_domain_index - 1;
178
179 if (!numa_enabled || index >= max_array_sz)
180 return NUMA_NO_NODE;
181
182 nid = of_read_number(&associativity[index], 1);
183
184 /* POWER4 LPAR uses 0xffff as invalid node */
185 if (nid == 0xffff || nid >= nr_node_ids)
186 nid = NUMA_NO_NODE;
187 return nid;
188}
189/*
190 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
191 * info is found.
192 */
193static int associativity_to_nid(const __be32 *associativity)
194{
195 int array_sz = of_read_number(associativity, 1);
196
197 /* Skip the first element in the associativity array */
198 return __associativity_to_nid((associativity + 1), array_sz);
199}
200
201static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
202{
203 int dist;
204 int node1, node2;
205
206 node1 = associativity_to_nid(cpu1_assoc);
207 node2 = associativity_to_nid(cpu2_assoc);
208
209 dist = numa_distance_table[node1][node2];
210 if (dist <= LOCAL_DISTANCE)
211 return 0;
212 else if (dist <= REMOTE_DISTANCE)
213 return 1;
214 else
215 return 2;
216}
217
218static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
219{
220 int dist = 0;
221
222 int i, index;
223
224 for (i = 0; i < distance_ref_points_depth; i++) {
225 index = be32_to_cpu(distance_ref_points[i]);
226 if (cpu1_assoc[index] == cpu2_assoc[index])
227 break;
228 dist++;
229 }
230
231 return dist;
232}
233
234int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
235{
236 /* We should not get called with FORM0 */
237 VM_WARN_ON(affinity_form == FORM0_AFFINITY);
238 if (affinity_form == FORM1_AFFINITY)
239 return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
240 return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
241}
242
243/* must hold reference to node during call */
244static const __be32 *of_get_associativity(struct device_node *dev)
245{
246 return of_get_property(dev, "ibm,associativity", NULL);
247}
248
249int __node_distance(int a, int b)
250{
251 int i;
252 int distance = LOCAL_DISTANCE;
253
254 if (affinity_form == FORM2_AFFINITY)
255 return numa_distance_table[a][b];
256 else if (affinity_form == FORM0_AFFINITY)
257 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
258
259 for (i = 0; i < distance_ref_points_depth; i++) {
260 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
261 break;
262
263 /* Double the distance for each NUMA level */
264 distance *= 2;
265 }
266
267 return distance;
268}
269EXPORT_SYMBOL(__node_distance);
270
271/* Returns the nid associated with the given device tree node,
272 * or -1 if not found.
273 */
274static int of_node_to_nid_single(struct device_node *device)
275{
276 int nid = NUMA_NO_NODE;
277 const __be32 *tmp;
278
279 tmp = of_get_associativity(device);
280 if (tmp)
281 nid = associativity_to_nid(tmp);
282 return nid;
283}
284
285/* Walk the device tree upwards, looking for an associativity id */
286int of_node_to_nid(struct device_node *device)
287{
288 int nid = NUMA_NO_NODE;
289
290 of_node_get(device);
291 while (device) {
292 nid = of_node_to_nid_single(device);
293 if (nid != -1)
294 break;
295
296 device = of_get_next_parent(device);
297 }
298 of_node_put(device);
299
300 return nid;
301}
302EXPORT_SYMBOL(of_node_to_nid);
303
304static void __initialize_form1_numa_distance(const __be32 *associativity,
305 int max_array_sz)
306{
307 int i, nid;
308
309 if (affinity_form != FORM1_AFFINITY)
310 return;
311
312 nid = __associativity_to_nid(associativity, max_array_sz);
313 if (nid != NUMA_NO_NODE) {
314 for (i = 0; i < distance_ref_points_depth; i++) {
315 const __be32 *entry;
316 int index = be32_to_cpu(distance_ref_points[i]) - 1;
317
318 /*
319 * broken hierarchy, return with broken distance table
320 */
321 if (WARN(index >= max_array_sz, "Broken ibm,associativity property"))
322 return;
323
324 entry = &associativity[index];
325 distance_lookup_table[nid][i] = of_read_number(entry, 1);
326 }
327 }
328}
329
330static void initialize_form1_numa_distance(const __be32 *associativity)
331{
332 int array_sz;
333
334 array_sz = of_read_number(associativity, 1);
335 /* Skip the first element in the associativity array */
336 __initialize_form1_numa_distance(associativity + 1, array_sz);
337}
338
339/*
340 * Used to update distance information w.r.t newly added node.
341 */
342void update_numa_distance(struct device_node *node)
343{
344 int nid;
345
346 if (affinity_form == FORM0_AFFINITY)
347 return;
348 else if (affinity_form == FORM1_AFFINITY) {
349 const __be32 *associativity;
350
351 associativity = of_get_associativity(node);
352 if (!associativity)
353 return;
354
355 initialize_form1_numa_distance(associativity);
356 return;
357 }
358
359 /* FORM2 affinity */
360 nid = of_node_to_nid_single(node);
361 if (nid == NUMA_NO_NODE)
362 return;
363
364 /*
365 * With FORM2 we expect NUMA distance of all possible NUMA
366 * nodes to be provided during boot.
367 */
368 WARN(numa_distance_table[nid][nid] == -1,
369 "NUMA distance details for node %d not provided\n", nid);
370}
371EXPORT_SYMBOL_GPL(update_numa_distance);
372
373/*
374 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
375 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
376 */
377static void __init initialize_form2_numa_distance_lookup_table(void)
378{
379 int i, j;
380 struct device_node *root;
381 const __u8 *form2_distances;
382 const __be32 *numa_lookup_index;
383 int form2_distances_length;
384 int max_numa_index, distance_index;
385
386 if (firmware_has_feature(FW_FEATURE_OPAL))
387 root = of_find_node_by_path("/ibm,opal");
388 else
389 root = of_find_node_by_path("/rtas");
390 if (!root)
391 root = of_find_node_by_path("/");
392
393 numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
394 max_numa_index = of_read_number(&numa_lookup_index[0], 1);
395
396 /* first element of the array is the size and is encode-int */
397 form2_distances = of_get_property(root, "ibm,numa-distance-table", NULL);
398 form2_distances_length = of_read_number((const __be32 *)&form2_distances[0], 1);
399 /* Skip the size which is encoded int */
400 form2_distances += sizeof(__be32);
401
402 pr_debug("form2_distances_len = %d, numa_dist_indexes_len = %d\n",
403 form2_distances_length, max_numa_index);
404
405 for (i = 0; i < max_numa_index; i++)
406 /* +1 skip the max_numa_index in the property */
407 numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
408
409
410 if (form2_distances_length != max_numa_index * max_numa_index) {
411 WARN(1, "Wrong NUMA distance information\n");
412 form2_distances = NULL; // don't use it
413 }
414 distance_index = 0;
415 for (i = 0; i < max_numa_index; i++) {
416 for (j = 0; j < max_numa_index; j++) {
417 int nodeA = numa_id_index_table[i];
418 int nodeB = numa_id_index_table[j];
419 int dist;
420
421 if (form2_distances)
422 dist = form2_distances[distance_index++];
423 else if (nodeA == nodeB)
424 dist = LOCAL_DISTANCE;
425 else
426 dist = REMOTE_DISTANCE;
427 numa_distance_table[nodeA][nodeB] = dist;
428 pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, dist);
429 }
430 }
431
432 of_node_put(root);
433}
434
435static int __init find_primary_domain_index(void)
436{
437 int index;
438 struct device_node *root;
439
440 /*
441 * Check for which form of affinity.
442 */
443 if (firmware_has_feature(FW_FEATURE_OPAL)) {
444 affinity_form = FORM1_AFFINITY;
445 } else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
446 pr_debug("Using form 2 affinity\n");
447 affinity_form = FORM2_AFFINITY;
448 } else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
449 pr_debug("Using form 1 affinity\n");
450 affinity_form = FORM1_AFFINITY;
451 } else
452 affinity_form = FORM0_AFFINITY;
453
454 if (firmware_has_feature(FW_FEATURE_OPAL))
455 root = of_find_node_by_path("/ibm,opal");
456 else
457 root = of_find_node_by_path("/rtas");
458 if (!root)
459 root = of_find_node_by_path("/");
460
461 /*
462 * This property is a set of 32-bit integers, each representing
463 * an index into the ibm,associativity nodes.
464 *
465 * With form 0 affinity the first integer is for an SMP configuration
466 * (should be all 0's) and the second is for a normal NUMA
467 * configuration. We have only one level of NUMA.
468 *
469 * With form 1 affinity the first integer is the most significant
470 * NUMA boundary and the following are progressively less significant
471 * boundaries. There can be more than one level of NUMA.
472 */
473 distance_ref_points = of_get_property(root,
474 "ibm,associativity-reference-points",
475 &distance_ref_points_depth);
476
477 if (!distance_ref_points) {
478 pr_debug("ibm,associativity-reference-points not found.\n");
479 goto err;
480 }
481
482 distance_ref_points_depth /= sizeof(int);
483 if (affinity_form == FORM0_AFFINITY) {
484 if (distance_ref_points_depth < 2) {
485 pr_warn("short ibm,associativity-reference-points\n");
486 goto err;
487 }
488
489 index = of_read_number(&distance_ref_points[1], 1);
490 } else {
491 /*
492 * Both FORM1 and FORM2 affinity find the primary domain details
493 * at the same offset.
494 */
495 index = of_read_number(distance_ref_points, 1);
496 }
497 /*
498 * Warn and cap if the hardware supports more than
499 * MAX_DISTANCE_REF_POINTS domains.
500 */
501 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
502 pr_warn("distance array capped at %d entries\n",
503 MAX_DISTANCE_REF_POINTS);
504 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
505 }
506
507 of_node_put(root);
508 return index;
509
510err:
511 of_node_put(root);
512 return -1;
513}
514
515static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
516{
517 struct device_node *memory = NULL;
518
519 memory = of_find_node_by_type(memory, "memory");
520 if (!memory)
521 panic("numa.c: No memory nodes found!");
522
523 *n_addr_cells = of_n_addr_cells(memory);
524 *n_size_cells = of_n_size_cells(memory);
525 of_node_put(memory);
526}
527
528static unsigned long read_n_cells(int n, const __be32 **buf)
529{
530 unsigned long result = 0;
531
532 while (n--) {
533 result = (result << 32) | of_read_number(*buf, 1);
534 (*buf)++;
535 }
536 return result;
537}
538
539struct assoc_arrays {
540 u32 n_arrays;
541 u32 array_sz;
542 const __be32 *arrays;
543};
544
545/*
546 * Retrieve and validate the list of associativity arrays for drconf
547 * memory from the ibm,associativity-lookup-arrays property of the
548 * device tree..
549 *
550 * The layout of the ibm,associativity-lookup-arrays property is a number N
551 * indicating the number of associativity arrays, followed by a number M
552 * indicating the size of each associativity array, followed by a list
553 * of N associativity arrays.
554 */
555static int of_get_assoc_arrays(struct assoc_arrays *aa)
556{
557 struct device_node *memory;
558 const __be32 *prop;
559 u32 len;
560
561 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
562 if (!memory)
563 return -1;
564
565 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
566 if (!prop || len < 2 * sizeof(unsigned int)) {
567 of_node_put(memory);
568 return -1;
569 }
570
571 aa->n_arrays = of_read_number(prop++, 1);
572 aa->array_sz = of_read_number(prop++, 1);
573
574 of_node_put(memory);
575
576 /* Now that we know the number of arrays and size of each array,
577 * revalidate the size of the property read in.
578 */
579 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
580 return -1;
581
582 aa->arrays = prop;
583 return 0;
584}
585
586static int __init get_nid_and_numa_distance(struct drmem_lmb *lmb)
587{
588 struct assoc_arrays aa = { .arrays = NULL };
589 int default_nid = NUMA_NO_NODE;
590 int nid = default_nid;
591 int rc, index;
592
593 if ((primary_domain_index < 0) || !numa_enabled)
594 return default_nid;
595
596 rc = of_get_assoc_arrays(&aa);
597 if (rc)
598 return default_nid;
599
600 if (primary_domain_index <= aa.array_sz &&
601 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
602 const __be32 *associativity;
603
604 index = lmb->aa_index * aa.array_sz;
605 associativity = &aa.arrays[index];
606 nid = __associativity_to_nid(associativity, aa.array_sz);
607 if (nid > 0 && affinity_form == FORM1_AFFINITY) {
608 /*
609 * lookup array associativity entries have
610 * no length of the array as the first element.
611 */
612 __initialize_form1_numa_distance(associativity, aa.array_sz);
613 }
614 }
615 return nid;
616}
617
618/*
619 * This is like of_node_to_nid_single() for memory represented in the
620 * ibm,dynamic-reconfiguration-memory node.
621 */
622int of_drconf_to_nid_single(struct drmem_lmb *lmb)
623{
624 struct assoc_arrays aa = { .arrays = NULL };
625 int default_nid = NUMA_NO_NODE;
626 int nid = default_nid;
627 int rc, index;
628
629 if ((primary_domain_index < 0) || !numa_enabled)
630 return default_nid;
631
632 rc = of_get_assoc_arrays(&aa);
633 if (rc)
634 return default_nid;
635
636 if (primary_domain_index <= aa.array_sz &&
637 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
638 const __be32 *associativity;
639
640 index = lmb->aa_index * aa.array_sz;
641 associativity = &aa.arrays[index];
642 nid = __associativity_to_nid(associativity, aa.array_sz);
643 }
644 return nid;
645}
646
647#ifdef CONFIG_PPC_SPLPAR
648
649static int __vphn_get_associativity(long lcpu, __be32 *associativity)
650{
651 long rc, hwid;
652
653 /*
654 * On a shared lpar, device tree will not have node associativity.
655 * At this time lppaca, or its __old_status field may not be
656 * updated. Hence kernel cannot detect if its on a shared lpar. So
657 * request an explicit associativity irrespective of whether the
658 * lpar is shared or dedicated. Use the device tree property as a
659 * fallback. cpu_to_phys_id is only valid between
660 * smp_setup_cpu_maps() and smp_setup_pacas().
661 */
662 if (firmware_has_feature(FW_FEATURE_VPHN)) {
663 if (cpu_to_phys_id)
664 hwid = cpu_to_phys_id[lcpu];
665 else
666 hwid = get_hard_smp_processor_id(lcpu);
667
668 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
669 if (rc == H_SUCCESS)
670 return 0;
671 }
672
673 return -1;
674}
675
676static int vphn_get_nid(long lcpu)
677{
678 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
679
680
681 if (!__vphn_get_associativity(lcpu, associativity))
682 return associativity_to_nid(associativity);
683
684 return NUMA_NO_NODE;
685
686}
687#else
688
689static int __vphn_get_associativity(long lcpu, __be32 *associativity)
690{
691 return -1;
692}
693
694static int vphn_get_nid(long unused)
695{
696 return NUMA_NO_NODE;
697}
698#endif /* CONFIG_PPC_SPLPAR */
699
700/*
701 * Figure out to which domain a cpu belongs and stick it there.
702 * Return the id of the domain used.
703 */
704static int numa_setup_cpu(unsigned long lcpu)
705{
706 struct device_node *cpu;
707 int fcpu = cpu_first_thread_sibling(lcpu);
708 int nid = NUMA_NO_NODE;
709
710 if (!cpu_present(lcpu)) {
711 set_cpu_numa_node(lcpu, first_online_node);
712 return first_online_node;
713 }
714
715 /*
716 * If a valid cpu-to-node mapping is already available, use it
717 * directly instead of querying the firmware, since it represents
718 * the most recent mapping notified to us by the platform (eg: VPHN).
719 * Since cpu_to_node binding remains the same for all threads in the
720 * core. If a valid cpu-to-node mapping is already available, for
721 * the first thread in the core, use it.
722 */
723 nid = numa_cpu_lookup_table[fcpu];
724 if (nid >= 0) {
725 map_cpu_to_node(lcpu, nid);
726 return nid;
727 }
728
729 nid = vphn_get_nid(lcpu);
730 if (nid != NUMA_NO_NODE)
731 goto out_present;
732
733 cpu = of_get_cpu_node(lcpu, NULL);
734
735 if (!cpu) {
736 WARN_ON(1);
737 if (cpu_present(lcpu))
738 goto out_present;
739 else
740 goto out;
741 }
742
743 nid = of_node_to_nid_single(cpu);
744 of_node_put(cpu);
745
746out_present:
747 if (nid < 0 || !node_possible(nid))
748 nid = first_online_node;
749
750 /*
751 * Update for the first thread of the core. All threads of a core
752 * have to be part of the same node. This not only avoids querying
753 * for every other thread in the core, but always avoids a case
754 * where virtual node associativity change causes subsequent threads
755 * of a core to be associated with different nid. However if first
756 * thread is already online, expect it to have a valid mapping.
757 */
758 if (fcpu != lcpu) {
759 WARN_ON(cpu_online(fcpu));
760 map_cpu_to_node(fcpu, nid);
761 }
762
763 map_cpu_to_node(lcpu, nid);
764out:
765 return nid;
766}
767
768static void verify_cpu_node_mapping(int cpu, int node)
769{
770 int base, sibling, i;
771
772 /* Verify that all the threads in the core belong to the same node */
773 base = cpu_first_thread_sibling(cpu);
774
775 for (i = 0; i < threads_per_core; i++) {
776 sibling = base + i;
777
778 if (sibling == cpu || cpu_is_offline(sibling))
779 continue;
780
781 if (cpu_to_node(sibling) != node) {
782 WARN(1, "CPU thread siblings %d and %d don't belong"
783 " to the same node!\n", cpu, sibling);
784 break;
785 }
786 }
787}
788
789/* Must run before sched domains notifier. */
790static int ppc_numa_cpu_prepare(unsigned int cpu)
791{
792 int nid;
793
794 nid = numa_setup_cpu(cpu);
795 verify_cpu_node_mapping(cpu, nid);
796 return 0;
797}
798
799static int ppc_numa_cpu_dead(unsigned int cpu)
800{
801 return 0;
802}
803
804/*
805 * Check and possibly modify a memory region to enforce the memory limit.
806 *
807 * Returns the size the region should have to enforce the memory limit.
808 * This will either be the original value of size, a truncated value,
809 * or zero. If the returned value of size is 0 the region should be
810 * discarded as it lies wholly above the memory limit.
811 */
812static unsigned long __init numa_enforce_memory_limit(unsigned long start,
813 unsigned long size)
814{
815 /*
816 * We use memblock_end_of_DRAM() in here instead of memory_limit because
817 * we've already adjusted it for the limit and it takes care of
818 * having memory holes below the limit. Also, in the case of
819 * iommu_is_off, memory_limit is not set but is implicitly enforced.
820 */
821
822 if (start + size <= memblock_end_of_DRAM())
823 return size;
824
825 if (start >= memblock_end_of_DRAM())
826 return 0;
827
828 return memblock_end_of_DRAM() - start;
829}
830
831/*
832 * Reads the counter for a given entry in
833 * linux,drconf-usable-memory property
834 */
835static inline int __init read_usm_ranges(const __be32 **usm)
836{
837 /*
838 * For each lmb in ibm,dynamic-memory a corresponding
839 * entry in linux,drconf-usable-memory property contains
840 * a counter followed by that many (base, size) duple.
841 * read the counter from linux,drconf-usable-memory
842 */
843 return read_n_cells(n_mem_size_cells, usm);
844}
845
846/*
847 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
848 * node. This assumes n_mem_{addr,size}_cells have been set.
849 */
850static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
851 const __be32 **usm,
852 void *data)
853{
854 unsigned int ranges, is_kexec_kdump = 0;
855 unsigned long base, size, sz;
856 int nid;
857
858 /*
859 * Skip this block if the reserved bit is set in flags (0x80)
860 * or if the block is not assigned to this partition (0x8)
861 */
862 if ((lmb->flags & DRCONF_MEM_RESERVED)
863 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
864 return 0;
865
866 if (*usm)
867 is_kexec_kdump = 1;
868
869 base = lmb->base_addr;
870 size = drmem_lmb_size();
871 ranges = 1;
872
873 if (is_kexec_kdump) {
874 ranges = read_usm_ranges(usm);
875 if (!ranges) /* there are no (base, size) duple */
876 return 0;
877 }
878
879 do {
880 if (is_kexec_kdump) {
881 base = read_n_cells(n_mem_addr_cells, usm);
882 size = read_n_cells(n_mem_size_cells, usm);
883 }
884
885 nid = get_nid_and_numa_distance(lmb);
886 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
887 &nid);
888 node_set_online(nid);
889 sz = numa_enforce_memory_limit(base, size);
890 if (sz)
891 memblock_set_node(base, sz, &memblock.memory, nid);
892 } while (--ranges);
893
894 return 0;
895}
896
897static int __init parse_numa_properties(void)
898{
899 struct device_node *memory;
900 int default_nid = 0;
901 unsigned long i;
902 const __be32 *associativity;
903
904 if (numa_enabled == 0) {
905 pr_warn("disabled by user\n");
906 return -1;
907 }
908
909 primary_domain_index = find_primary_domain_index();
910
911 if (primary_domain_index < 0) {
912 /*
913 * if we fail to parse primary_domain_index from device tree
914 * mark the numa disabled, boot with numa disabled.
915 */
916 numa_enabled = false;
917 return primary_domain_index;
918 }
919
920 pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index);
921
922 /*
923 * If it is FORM2 initialize the distance table here.
924 */
925 if (affinity_form == FORM2_AFFINITY)
926 initialize_form2_numa_distance_lookup_table();
927
928 /*
929 * Even though we connect cpus to numa domains later in SMP
930 * init, we need to know the node ids now. This is because
931 * each node to be onlined must have NODE_DATA etc backing it.
932 */
933 for_each_present_cpu(i) {
934 __be32 vphn_assoc[VPHN_ASSOC_BUFSIZE];
935 struct device_node *cpu;
936 int nid = NUMA_NO_NODE;
937
938 memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32));
939
940 if (__vphn_get_associativity(i, vphn_assoc) == 0) {
941 nid = associativity_to_nid(vphn_assoc);
942 initialize_form1_numa_distance(vphn_assoc);
943 } else {
944
945 /*
946 * Don't fall back to default_nid yet -- we will plug
947 * cpus into nodes once the memory scan has discovered
948 * the topology.
949 */
950 cpu = of_get_cpu_node(i, NULL);
951 BUG_ON(!cpu);
952
953 associativity = of_get_associativity(cpu);
954 if (associativity) {
955 nid = associativity_to_nid(associativity);
956 initialize_form1_numa_distance(associativity);
957 }
958 of_node_put(cpu);
959 }
960
961 /* node_set_online() is an UB if 'nid' is negative */
962 if (likely(nid >= 0))
963 node_set_online(nid);
964 }
965
966 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
967
968 for_each_node_by_type(memory, "memory") {
969 unsigned long start;
970 unsigned long size;
971 int nid;
972 int ranges;
973 const __be32 *memcell_buf;
974 unsigned int len;
975
976 memcell_buf = of_get_property(memory,
977 "linux,usable-memory", &len);
978 if (!memcell_buf || len <= 0)
979 memcell_buf = of_get_property(memory, "reg", &len);
980 if (!memcell_buf || len <= 0)
981 continue;
982
983 /* ranges in cell */
984 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
985new_range:
986 /* these are order-sensitive, and modify the buffer pointer */
987 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
988 size = read_n_cells(n_mem_size_cells, &memcell_buf);
989
990 /*
991 * Assumption: either all memory nodes or none will
992 * have associativity properties. If none, then
993 * everything goes to default_nid.
994 */
995 associativity = of_get_associativity(memory);
996 if (associativity) {
997 nid = associativity_to_nid(associativity);
998 initialize_form1_numa_distance(associativity);
999 } else
1000 nid = default_nid;
1001
1002 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
1003 node_set_online(nid);
1004
1005 size = numa_enforce_memory_limit(start, size);
1006 if (size)
1007 memblock_set_node(start, size, &memblock.memory, nid);
1008
1009 if (--ranges)
1010 goto new_range;
1011 }
1012
1013 /*
1014 * Now do the same thing for each MEMBLOCK listed in the
1015 * ibm,dynamic-memory property in the
1016 * ibm,dynamic-reconfiguration-memory node.
1017 */
1018 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1019 if (memory) {
1020 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
1021 of_node_put(memory);
1022 }
1023
1024 return 0;
1025}
1026
1027static void __init setup_nonnuma(void)
1028{
1029 unsigned long top_of_ram = memblock_end_of_DRAM();
1030 unsigned long total_ram = memblock_phys_mem_size();
1031 unsigned long start_pfn, end_pfn;
1032 unsigned int nid = 0;
1033 int i;
1034
1035 pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram);
1036 pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20);
1037
1038 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1039 fake_numa_create_new_node(end_pfn, &nid);
1040 memblock_set_node(PFN_PHYS(start_pfn),
1041 PFN_PHYS(end_pfn - start_pfn),
1042 &memblock.memory, nid);
1043 node_set_online(nid);
1044 }
1045}
1046
1047void __init dump_numa_cpu_topology(void)
1048{
1049 unsigned int node;
1050 unsigned int cpu, count;
1051
1052 if (!numa_enabled)
1053 return;
1054
1055 for_each_online_node(node) {
1056 pr_info("Node %d CPUs:", node);
1057
1058 count = 0;
1059 /*
1060 * If we used a CPU iterator here we would miss printing
1061 * the holes in the cpumap.
1062 */
1063 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1064 if (cpumask_test_cpu(cpu,
1065 node_to_cpumask_map[node])) {
1066 if (count == 0)
1067 pr_cont(" %u", cpu);
1068 ++count;
1069 } else {
1070 if (count > 1)
1071 pr_cont("-%u", cpu - 1);
1072 count = 0;
1073 }
1074 }
1075
1076 if (count > 1)
1077 pr_cont("-%u", nr_cpu_ids - 1);
1078 pr_cont("\n");
1079 }
1080}
1081
1082/* Initialize NODE_DATA for a node on the local memory */
1083static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
1084{
1085 u64 spanned_pages = end_pfn - start_pfn;
1086 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
1087 u64 nd_pa;
1088 void *nd;
1089 int tnid;
1090
1091 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
1092 if (!nd_pa)
1093 panic("Cannot allocate %zu bytes for node %d data\n",
1094 nd_size, nid);
1095
1096 nd = __va(nd_pa);
1097
1098 /* report and initialize */
1099 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
1100 nd_pa, nd_pa + nd_size - 1);
1101 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
1102 if (tnid != nid)
1103 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
1104
1105 node_data[nid] = nd;
1106 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
1107 NODE_DATA(nid)->node_id = nid;
1108 NODE_DATA(nid)->node_start_pfn = start_pfn;
1109 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
1110}
1111
1112static void __init find_possible_nodes(void)
1113{
1114 struct device_node *rtas;
1115 const __be32 *domains = NULL;
1116 int prop_length, max_nodes;
1117 u32 i;
1118
1119 if (!numa_enabled)
1120 return;
1121
1122 rtas = of_find_node_by_path("/rtas");
1123 if (!rtas)
1124 return;
1125
1126 /*
1127 * ibm,current-associativity-domains is a fairly recent property. If
1128 * it doesn't exist, then fallback on ibm,max-associativity-domains.
1129 * Current denotes what the platform can support compared to max
1130 * which denotes what the Hypervisor can support.
1131 *
1132 * If the LPAR is migratable, new nodes might be activated after a LPM,
1133 * so we should consider the max number in that case.
1134 */
1135 if (!of_get_property(of_root, "ibm,migratable-partition", NULL))
1136 domains = of_get_property(rtas,
1137 "ibm,current-associativity-domains",
1138 &prop_length);
1139 if (!domains) {
1140 domains = of_get_property(rtas, "ibm,max-associativity-domains",
1141 &prop_length);
1142 if (!domains)
1143 goto out;
1144 }
1145
1146 max_nodes = of_read_number(&domains[primary_domain_index], 1);
1147 pr_info("Partition configured for %d NUMA nodes.\n", max_nodes);
1148
1149 for (i = 0; i < max_nodes; i++) {
1150 if (!node_possible(i))
1151 node_set(i, node_possible_map);
1152 }
1153
1154 prop_length /= sizeof(int);
1155 if (prop_length > primary_domain_index + 2)
1156 coregroup_enabled = 1;
1157
1158out:
1159 of_node_put(rtas);
1160}
1161
1162void __init mem_topology_setup(void)
1163{
1164 int cpu;
1165
1166 max_low_pfn = max_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1167 min_low_pfn = MEMORY_START >> PAGE_SHIFT;
1168
1169 /*
1170 * Linux/mm assumes node 0 to be online at boot. However this is not
1171 * true on PowerPC, where node 0 is similar to any other node, it
1172 * could be cpuless, memoryless node. So force node 0 to be offline
1173 * for now. This will prevent cpuless, memoryless node 0 showing up
1174 * unnecessarily as online. If a node has cpus or memory that need
1175 * to be online, then node will anyway be marked online.
1176 */
1177 node_set_offline(0);
1178
1179 if (parse_numa_properties())
1180 setup_nonnuma();
1181
1182 /*
1183 * Modify the set of possible NUMA nodes to reflect information
1184 * available about the set of online nodes, and the set of nodes
1185 * that we expect to make use of for this platform's affinity
1186 * calculations.
1187 */
1188 nodes_and(node_possible_map, node_possible_map, node_online_map);
1189
1190 find_possible_nodes();
1191
1192 setup_node_to_cpumask_map();
1193
1194 reset_numa_cpu_lookup_table();
1195
1196 for_each_possible_cpu(cpu) {
1197 /*
1198 * Powerpc with CONFIG_NUMA always used to have a node 0,
1199 * even if it was memoryless or cpuless. For all cpus that
1200 * are possible but not present, cpu_to_node() would point
1201 * to node 0. To remove a cpuless, memoryless dummy node,
1202 * powerpc need to make sure all possible but not present
1203 * cpu_to_node are set to a proper node.
1204 */
1205 numa_setup_cpu(cpu);
1206 }
1207}
1208
1209void __init initmem_init(void)
1210{
1211 int nid;
1212
1213 memblock_dump_all();
1214
1215 for_each_online_node(nid) {
1216 unsigned long start_pfn, end_pfn;
1217
1218 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1219 setup_node_data(nid, start_pfn, end_pfn);
1220 }
1221
1222 sparse_init();
1223
1224 /*
1225 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1226 * even before we online them, so that we can use cpu_to_{node,mem}
1227 * early in boot, cf. smp_prepare_cpus().
1228 * _nocalls() + manual invocation is used because cpuhp is not yet
1229 * initialized for the boot CPU.
1230 */
1231 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1232 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
1233}
1234
1235static int __init early_numa(char *p)
1236{
1237 if (!p)
1238 return 0;
1239
1240 if (strstr(p, "off"))
1241 numa_enabled = 0;
1242
1243 p = strstr(p, "fake=");
1244 if (p)
1245 cmdline = p + strlen("fake=");
1246
1247 return 0;
1248}
1249early_param("numa", early_numa);
1250
1251#ifdef CONFIG_MEMORY_HOTPLUG
1252/*
1253 * Find the node associated with a hot added memory section for
1254 * memory represented in the device tree by the property
1255 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1256 */
1257static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
1258{
1259 struct drmem_lmb *lmb;
1260 unsigned long lmb_size;
1261 int nid = NUMA_NO_NODE;
1262
1263 lmb_size = drmem_lmb_size();
1264
1265 for_each_drmem_lmb(lmb) {
1266 /* skip this block if it is reserved or not assigned to
1267 * this partition */
1268 if ((lmb->flags & DRCONF_MEM_RESERVED)
1269 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1270 continue;
1271
1272 if ((scn_addr < lmb->base_addr)
1273 || (scn_addr >= (lmb->base_addr + lmb_size)))
1274 continue;
1275
1276 nid = of_drconf_to_nid_single(lmb);
1277 break;
1278 }
1279
1280 return nid;
1281}
1282
1283/*
1284 * Find the node associated with a hot added memory section for memory
1285 * represented in the device tree as a node (i.e. memory@XXXX) for
1286 * each memblock.
1287 */
1288static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1289{
1290 struct device_node *memory;
1291 int nid = NUMA_NO_NODE;
1292
1293 for_each_node_by_type(memory, "memory") {
1294 int i = 0;
1295
1296 while (1) {
1297 struct resource res;
1298
1299 if (of_address_to_resource(memory, i++, &res))
1300 break;
1301
1302 if ((scn_addr < res.start) || (scn_addr > res.end))
1303 continue;
1304
1305 nid = of_node_to_nid_single(memory);
1306 break;
1307 }
1308
1309 if (nid >= 0)
1310 break;
1311 }
1312
1313 of_node_put(memory);
1314
1315 return nid;
1316}
1317
1318/*
1319 * Find the node associated with a hot added memory section. Section
1320 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1321 * sections are fully contained within a single MEMBLOCK.
1322 */
1323int hot_add_scn_to_nid(unsigned long scn_addr)
1324{
1325 struct device_node *memory = NULL;
1326 int nid;
1327
1328 if (!numa_enabled)
1329 return first_online_node;
1330
1331 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1332 if (memory) {
1333 nid = hot_add_drconf_scn_to_nid(scn_addr);
1334 of_node_put(memory);
1335 } else {
1336 nid = hot_add_node_scn_to_nid(scn_addr);
1337 }
1338
1339 if (nid < 0 || !node_possible(nid))
1340 nid = first_online_node;
1341
1342 return nid;
1343}
1344
1345static u64 hot_add_drconf_memory_max(void)
1346{
1347 struct device_node *memory = NULL;
1348 struct device_node *dn = NULL;
1349 const __be64 *lrdr = NULL;
1350
1351 dn = of_find_node_by_path("/rtas");
1352 if (dn) {
1353 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1354 of_node_put(dn);
1355 if (lrdr)
1356 return be64_to_cpup(lrdr);
1357 }
1358
1359 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1360 if (memory) {
1361 of_node_put(memory);
1362 return drmem_lmb_memory_max();
1363 }
1364 return 0;
1365}
1366
1367/*
1368 * memory_hotplug_max - return max address of memory that may be added
1369 *
1370 * This is currently only used on systems that support drconfig memory
1371 * hotplug.
1372 */
1373u64 memory_hotplug_max(void)
1374{
1375 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1376}
1377#endif /* CONFIG_MEMORY_HOTPLUG */
1378
1379/* Virtual Processor Home Node (VPHN) support */
1380#ifdef CONFIG_PPC_SPLPAR
1381static int topology_inited;
1382
1383/*
1384 * Retrieve the new associativity information for a virtual processor's
1385 * home node.
1386 */
1387static long vphn_get_associativity(unsigned long cpu,
1388 __be32 *associativity)
1389{
1390 long rc;
1391
1392 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1393 VPHN_FLAG_VCPU, associativity);
1394
1395 switch (rc) {
1396 case H_SUCCESS:
1397 pr_debug("VPHN hcall succeeded. Reset polling...\n");
1398 goto out;
1399
1400 case H_FUNCTION:
1401 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
1402 break;
1403 case H_HARDWARE:
1404 pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
1405 "preventing VPHN. Disabling polling...\n");
1406 break;
1407 case H_PARAMETER:
1408 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1409 "Disabling polling...\n");
1410 break;
1411 default:
1412 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1413 , rc);
1414 break;
1415 }
1416out:
1417 return rc;
1418}
1419
1420void find_and_update_cpu_nid(int cpu)
1421{
1422 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1423 int new_nid;
1424
1425 /* Use associativity from first thread for all siblings */
1426 if (vphn_get_associativity(cpu, associativity))
1427 return;
1428
1429 /* Do not have previous associativity, so find it now. */
1430 new_nid = associativity_to_nid(associativity);
1431
1432 if (new_nid < 0 || !node_possible(new_nid))
1433 new_nid = first_online_node;
1434 else
1435 // Associate node <-> cpu, so cpu_up() calls
1436 // try_online_node() on the right node.
1437 set_cpu_numa_node(cpu, new_nid);
1438
1439 pr_debug("%s:%d cpu %d nid %d\n", __func__, __LINE__, cpu, new_nid);
1440}
1441
1442int cpu_to_coregroup_id(int cpu)
1443{
1444 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1445 int index;
1446
1447 if (cpu < 0 || cpu > nr_cpu_ids)
1448 return -1;
1449
1450 if (!coregroup_enabled)
1451 goto out;
1452
1453 if (!firmware_has_feature(FW_FEATURE_VPHN))
1454 goto out;
1455
1456 if (vphn_get_associativity(cpu, associativity))
1457 goto out;
1458
1459 index = of_read_number(associativity, 1);
1460 if (index > primary_domain_index + 1)
1461 return of_read_number(&associativity[index - 1], 1);
1462
1463out:
1464 return cpu_to_core_id(cpu);
1465}
1466
1467static int topology_update_init(void)
1468{
1469 topology_inited = 1;
1470 return 0;
1471}
1472device_initcall(topology_update_init);
1473#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/export.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 <linux/stop_machine.h>
26#include <linux/proc_fs.h>
27#include <linux/seq_file.h>
28#include <linux/uaccess.h>
29#include <linux/slab.h>
30#include <asm/cputhreads.h>
31#include <asm/sparsemem.h>
32#include <asm/prom.h>
33#include <asm/smp.h>
34#include <asm/cputhreads.h>
35#include <asm/topology.h>
36#include <asm/firmware.h>
37#include <asm/paca.h>
38#include <asm/hvcall.h>
39#include <asm/setup.h>
40#include <asm/vdso.h>
41
42static int numa_enabled = 1;
43
44static char *cmdline __initdata;
45
46static int numa_debug;
47#define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
48
49int numa_cpu_lookup_table[NR_CPUS];
50cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
51struct pglist_data *node_data[MAX_NUMNODES];
52
53EXPORT_SYMBOL(numa_cpu_lookup_table);
54EXPORT_SYMBOL(node_to_cpumask_map);
55EXPORT_SYMBOL(node_data);
56
57static int min_common_depth;
58static int n_mem_addr_cells, n_mem_size_cells;
59static int form1_affinity;
60
61#define MAX_DISTANCE_REF_POINTS 4
62static int distance_ref_points_depth;
63static const __be32 *distance_ref_points;
64static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
65
66/*
67 * Allocate node_to_cpumask_map based on number of available nodes
68 * Requires node_possible_map to be valid.
69 *
70 * Note: cpumask_of_node() is not valid until after this is done.
71 */
72static void __init setup_node_to_cpumask_map(void)
73{
74 unsigned int node;
75
76 /* setup nr_node_ids if not done yet */
77 if (nr_node_ids == MAX_NUMNODES)
78 setup_nr_node_ids();
79
80 /* allocate the map */
81 for (node = 0; node < nr_node_ids; node++)
82 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
83
84 /* cpumask_of_node() will now work */
85 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
86}
87
88static int __init fake_numa_create_new_node(unsigned long end_pfn,
89 unsigned int *nid)
90{
91 unsigned long long mem;
92 char *p = cmdline;
93 static unsigned int fake_nid;
94 static unsigned long long curr_boundary;
95
96 /*
97 * Modify node id, iff we started creating NUMA nodes
98 * We want to continue from where we left of the last time
99 */
100 if (fake_nid)
101 *nid = fake_nid;
102 /*
103 * In case there are no more arguments to parse, the
104 * node_id should be the same as the last fake node id
105 * (we've handled this above).
106 */
107 if (!p)
108 return 0;
109
110 mem = memparse(p, &p);
111 if (!mem)
112 return 0;
113
114 if (mem < curr_boundary)
115 return 0;
116
117 curr_boundary = mem;
118
119 if ((end_pfn << PAGE_SHIFT) > mem) {
120 /*
121 * Skip commas and spaces
122 */
123 while (*p == ',' || *p == ' ' || *p == '\t')
124 p++;
125
126 cmdline = p;
127 fake_nid++;
128 *nid = fake_nid;
129 dbg("created new fake_node with id %d\n", fake_nid);
130 return 1;
131 }
132 return 0;
133}
134
135/*
136 * get_node_active_region - Return active region containing pfn
137 * Active range returned is empty if none found.
138 * @pfn: The page to return the region for
139 * @node_ar: Returned set to the active region containing @pfn
140 */
141static void __init get_node_active_region(unsigned long pfn,
142 struct node_active_region *node_ar)
143{
144 unsigned long start_pfn, end_pfn;
145 int i, nid;
146
147 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
148 if (pfn >= start_pfn && pfn < end_pfn) {
149 node_ar->nid = nid;
150 node_ar->start_pfn = start_pfn;
151 node_ar->end_pfn = end_pfn;
152 break;
153 }
154 }
155}
156
157static void reset_numa_cpu_lookup_table(void)
158{
159 unsigned int cpu;
160
161 for_each_possible_cpu(cpu)
162 numa_cpu_lookup_table[cpu] = -1;
163}
164
165static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
166{
167 numa_cpu_lookup_table[cpu] = node;
168}
169
170static void map_cpu_to_node(int cpu, int node)
171{
172 update_numa_cpu_lookup_table(cpu, node);
173
174 dbg("adding cpu %d to node %d\n", cpu, node);
175
176 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
177 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
178}
179
180#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
181static void unmap_cpu_from_node(unsigned long cpu)
182{
183 int node = numa_cpu_lookup_table[cpu];
184
185 dbg("removing cpu %lu from node %d\n", cpu, node);
186
187 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
188 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
189 } else {
190 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
191 cpu, node);
192 }
193}
194#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
195
196/* must hold reference to node during call */
197static const __be32 *of_get_associativity(struct device_node *dev)
198{
199 return of_get_property(dev, "ibm,associativity", NULL);
200}
201
202/*
203 * Returns the property linux,drconf-usable-memory if
204 * it exists (the property exists only in kexec/kdump kernels,
205 * added by kexec-tools)
206 */
207static const __be32 *of_get_usable_memory(struct device_node *memory)
208{
209 const __be32 *prop;
210 u32 len;
211 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
212 if (!prop || len < sizeof(unsigned int))
213 return NULL;
214 return prop;
215}
216
217int __node_distance(int a, int b)
218{
219 int i;
220 int distance = LOCAL_DISTANCE;
221
222 if (!form1_affinity)
223 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
224
225 for (i = 0; i < distance_ref_points_depth; i++) {
226 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
227 break;
228
229 /* Double the distance for each NUMA level */
230 distance *= 2;
231 }
232
233 return distance;
234}
235EXPORT_SYMBOL(__node_distance);
236
237static void initialize_distance_lookup_table(int nid,
238 const __be32 *associativity)
239{
240 int i;
241
242 if (!form1_affinity)
243 return;
244
245 for (i = 0; i < distance_ref_points_depth; i++) {
246 const __be32 *entry;
247
248 entry = &associativity[be32_to_cpu(distance_ref_points[i])];
249 distance_lookup_table[nid][i] = of_read_number(entry, 1);
250 }
251}
252
253/* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
254 * info is found.
255 */
256static int associativity_to_nid(const __be32 *associativity)
257{
258 int nid = -1;
259
260 if (min_common_depth == -1)
261 goto out;
262
263 if (of_read_number(associativity, 1) >= min_common_depth)
264 nid = of_read_number(&associativity[min_common_depth], 1);
265
266 /* POWER4 LPAR uses 0xffff as invalid node */
267 if (nid == 0xffff || nid >= MAX_NUMNODES)
268 nid = -1;
269
270 if (nid > 0 &&
271 of_read_number(associativity, 1) >= distance_ref_points_depth)
272 initialize_distance_lookup_table(nid, associativity);
273
274out:
275 return nid;
276}
277
278/* Returns the nid associated with the given device tree node,
279 * or -1 if not found.
280 */
281static int of_node_to_nid_single(struct device_node *device)
282{
283 int nid = -1;
284 const __be32 *tmp;
285
286 tmp = of_get_associativity(device);
287 if (tmp)
288 nid = associativity_to_nid(tmp);
289 return nid;
290}
291
292/* Walk the device tree upwards, looking for an associativity id */
293int of_node_to_nid(struct device_node *device)
294{
295 struct device_node *tmp;
296 int nid = -1;
297
298 of_node_get(device);
299 while (device) {
300 nid = of_node_to_nid_single(device);
301 if (nid != -1)
302 break;
303
304 tmp = device;
305 device = of_get_parent(tmp);
306 of_node_put(tmp);
307 }
308 of_node_put(device);
309
310 return nid;
311}
312EXPORT_SYMBOL_GPL(of_node_to_nid);
313
314static int __init find_min_common_depth(void)
315{
316 int depth;
317 struct device_node *root;
318
319 if (firmware_has_feature(FW_FEATURE_OPAL))
320 root = of_find_node_by_path("/ibm,opal");
321 else
322 root = of_find_node_by_path("/rtas");
323 if (!root)
324 root = of_find_node_by_path("/");
325
326 /*
327 * This property is a set of 32-bit integers, each representing
328 * an index into the ibm,associativity nodes.
329 *
330 * With form 0 affinity the first integer is for an SMP configuration
331 * (should be all 0's) and the second is for a normal NUMA
332 * configuration. We have only one level of NUMA.
333 *
334 * With form 1 affinity the first integer is the most significant
335 * NUMA boundary and the following are progressively less significant
336 * boundaries. There can be more than one level of NUMA.
337 */
338 distance_ref_points = of_get_property(root,
339 "ibm,associativity-reference-points",
340 &distance_ref_points_depth);
341
342 if (!distance_ref_points) {
343 dbg("NUMA: ibm,associativity-reference-points not found.\n");
344 goto err;
345 }
346
347 distance_ref_points_depth /= sizeof(int);
348
349 if (firmware_has_feature(FW_FEATURE_OPAL) ||
350 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
351 dbg("Using form 1 affinity\n");
352 form1_affinity = 1;
353 }
354
355 if (form1_affinity) {
356 depth = of_read_number(distance_ref_points, 1);
357 } else {
358 if (distance_ref_points_depth < 2) {
359 printk(KERN_WARNING "NUMA: "
360 "short ibm,associativity-reference-points\n");
361 goto err;
362 }
363
364 depth = of_read_number(&distance_ref_points[1], 1);
365 }
366
367 /*
368 * Warn and cap if the hardware supports more than
369 * MAX_DISTANCE_REF_POINTS domains.
370 */
371 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
372 printk(KERN_WARNING "NUMA: distance array capped at "
373 "%d entries\n", MAX_DISTANCE_REF_POINTS);
374 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
375 }
376
377 of_node_put(root);
378 return depth;
379
380err:
381 of_node_put(root);
382 return -1;
383}
384
385static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
386{
387 struct device_node *memory = NULL;
388
389 memory = of_find_node_by_type(memory, "memory");
390 if (!memory)
391 panic("numa.c: No memory nodes found!");
392
393 *n_addr_cells = of_n_addr_cells(memory);
394 *n_size_cells = of_n_size_cells(memory);
395 of_node_put(memory);
396}
397
398static unsigned long read_n_cells(int n, const __be32 **buf)
399{
400 unsigned long result = 0;
401
402 while (n--) {
403 result = (result << 32) | of_read_number(*buf, 1);
404 (*buf)++;
405 }
406 return result;
407}
408
409/*
410 * Read the next memblock list entry from the ibm,dynamic-memory property
411 * and return the information in the provided of_drconf_cell structure.
412 */
413static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
414{
415 const __be32 *cp;
416
417 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
418
419 cp = *cellp;
420 drmem->drc_index = of_read_number(cp, 1);
421 drmem->reserved = of_read_number(&cp[1], 1);
422 drmem->aa_index = of_read_number(&cp[2], 1);
423 drmem->flags = of_read_number(&cp[3], 1);
424
425 *cellp = cp + 4;
426}
427
428/*
429 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
430 *
431 * The layout of the ibm,dynamic-memory property is a number N of memblock
432 * list entries followed by N memblock list entries. Each memblock list entry
433 * contains information as laid out in the of_drconf_cell struct above.
434 */
435static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
436{
437 const __be32 *prop;
438 u32 len, entries;
439
440 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
441 if (!prop || len < sizeof(unsigned int))
442 return 0;
443
444 entries = of_read_number(prop++, 1);
445
446 /* Now that we know the number of entries, revalidate the size
447 * of the property read in to ensure we have everything
448 */
449 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
450 return 0;
451
452 *dm = prop;
453 return entries;
454}
455
456/*
457 * Retrieve and validate the ibm,lmb-size property for drconf memory
458 * from the device tree.
459 */
460static u64 of_get_lmb_size(struct device_node *memory)
461{
462 const __be32 *prop;
463 u32 len;
464
465 prop = of_get_property(memory, "ibm,lmb-size", &len);
466 if (!prop || len < sizeof(unsigned int))
467 return 0;
468
469 return read_n_cells(n_mem_size_cells, &prop);
470}
471
472struct assoc_arrays {
473 u32 n_arrays;
474 u32 array_sz;
475 const __be32 *arrays;
476};
477
478/*
479 * Retrieve and validate the list of associativity arrays for drconf
480 * memory from the ibm,associativity-lookup-arrays property of the
481 * device tree..
482 *
483 * The layout of the ibm,associativity-lookup-arrays property is a number N
484 * indicating the number of associativity arrays, followed by a number M
485 * indicating the size of each associativity array, followed by a list
486 * of N associativity arrays.
487 */
488static int of_get_assoc_arrays(struct device_node *memory,
489 struct assoc_arrays *aa)
490{
491 const __be32 *prop;
492 u32 len;
493
494 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
495 if (!prop || len < 2 * sizeof(unsigned int))
496 return -1;
497
498 aa->n_arrays = of_read_number(prop++, 1);
499 aa->array_sz = of_read_number(prop++, 1);
500
501 /* Now that we know the number of arrays and size of each array,
502 * revalidate the size of the property read in.
503 */
504 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
505 return -1;
506
507 aa->arrays = prop;
508 return 0;
509}
510
511/*
512 * This is like of_node_to_nid_single() for memory represented in the
513 * ibm,dynamic-reconfiguration-memory node.
514 */
515static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
516 struct assoc_arrays *aa)
517{
518 int default_nid = 0;
519 int nid = default_nid;
520 int index;
521
522 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
523 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
524 drmem->aa_index < aa->n_arrays) {
525 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
526 nid = of_read_number(&aa->arrays[index], 1);
527
528 if (nid == 0xffff || nid >= MAX_NUMNODES)
529 nid = default_nid;
530 }
531
532 return nid;
533}
534
535/*
536 * Figure out to which domain a cpu belongs and stick it there.
537 * Return the id of the domain used.
538 */
539static int numa_setup_cpu(unsigned long lcpu)
540{
541 int nid;
542 struct device_node *cpu;
543
544 /*
545 * If a valid cpu-to-node mapping is already available, use it
546 * directly instead of querying the firmware, since it represents
547 * the most recent mapping notified to us by the platform (eg: VPHN).
548 */
549 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
550 map_cpu_to_node(lcpu, nid);
551 return nid;
552 }
553
554 cpu = of_get_cpu_node(lcpu, NULL);
555
556 if (!cpu) {
557 WARN_ON(1);
558 nid = 0;
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 void verify_cpu_node_mapping(int cpu, int node)
575{
576 int base, sibling, i;
577
578 /* Verify that all the threads in the core belong to the same node */
579 base = cpu_first_thread_sibling(cpu);
580
581 for (i = 0; i < threads_per_core; i++) {
582 sibling = base + i;
583
584 if (sibling == cpu || cpu_is_offline(sibling))
585 continue;
586
587 if (cpu_to_node(sibling) != node) {
588 WARN(1, "CPU thread siblings %d and %d don't belong"
589 " to the same node!\n", cpu, sibling);
590 break;
591 }
592 }
593}
594
595static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
596 void *hcpu)
597{
598 unsigned long lcpu = (unsigned long)hcpu;
599 int ret = NOTIFY_DONE, nid;
600
601 switch (action) {
602 case CPU_UP_PREPARE:
603 case CPU_UP_PREPARE_FROZEN:
604 nid = numa_setup_cpu(lcpu);
605 verify_cpu_node_mapping((int)lcpu, nid);
606 ret = NOTIFY_OK;
607 break;
608#ifdef CONFIG_HOTPLUG_CPU
609 case CPU_DEAD:
610 case CPU_DEAD_FROZEN:
611 case CPU_UP_CANCELED:
612 case CPU_UP_CANCELED_FROZEN:
613 unmap_cpu_from_node(lcpu);
614 break;
615 ret = NOTIFY_OK;
616#endif
617 }
618 return ret;
619}
620
621/*
622 * Check and possibly modify a memory region to enforce the memory limit.
623 *
624 * Returns the size the region should have to enforce the memory limit.
625 * This will either be the original value of size, a truncated value,
626 * or zero. If the returned value of size is 0 the region should be
627 * discarded as it lies wholly above the memory limit.
628 */
629static unsigned long __init numa_enforce_memory_limit(unsigned long start,
630 unsigned long size)
631{
632 /*
633 * We use memblock_end_of_DRAM() in here instead of memory_limit because
634 * we've already adjusted it for the limit and it takes care of
635 * having memory holes below the limit. Also, in the case of
636 * iommu_is_off, memory_limit is not set but is implicitly enforced.
637 */
638
639 if (start + size <= memblock_end_of_DRAM())
640 return size;
641
642 if (start >= memblock_end_of_DRAM())
643 return 0;
644
645 return memblock_end_of_DRAM() - start;
646}
647
648/*
649 * Reads the counter for a given entry in
650 * linux,drconf-usable-memory property
651 */
652static inline int __init read_usm_ranges(const __be32 **usm)
653{
654 /*
655 * For each lmb in ibm,dynamic-memory a corresponding
656 * entry in linux,drconf-usable-memory property contains
657 * a counter followed by that many (base, size) duple.
658 * read the counter from linux,drconf-usable-memory
659 */
660 return read_n_cells(n_mem_size_cells, usm);
661}
662
663/*
664 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
665 * node. This assumes n_mem_{addr,size}_cells have been set.
666 */
667static void __init parse_drconf_memory(struct device_node *memory)
668{
669 const __be32 *uninitialized_var(dm), *usm;
670 unsigned int n, rc, ranges, is_kexec_kdump = 0;
671 unsigned long lmb_size, base, size, sz;
672 int nid;
673 struct assoc_arrays aa = { .arrays = NULL };
674
675 n = of_get_drconf_memory(memory, &dm);
676 if (!n)
677 return;
678
679 lmb_size = of_get_lmb_size(memory);
680 if (!lmb_size)
681 return;
682
683 rc = of_get_assoc_arrays(memory, &aa);
684 if (rc)
685 return;
686
687 /* check if this is a kexec/kdump kernel */
688 usm = of_get_usable_memory(memory);
689 if (usm != NULL)
690 is_kexec_kdump = 1;
691
692 for (; n != 0; --n) {
693 struct of_drconf_cell drmem;
694
695 read_drconf_cell(&drmem, &dm);
696
697 /* skip this block if the reserved bit is set in flags (0x80)
698 or if the block is not assigned to this partition (0x8) */
699 if ((drmem.flags & DRCONF_MEM_RESERVED)
700 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
701 continue;
702
703 base = drmem.base_addr;
704 size = lmb_size;
705 ranges = 1;
706
707 if (is_kexec_kdump) {
708 ranges = read_usm_ranges(&usm);
709 if (!ranges) /* there are no (base, size) duple */
710 continue;
711 }
712 do {
713 if (is_kexec_kdump) {
714 base = read_n_cells(n_mem_addr_cells, &usm);
715 size = read_n_cells(n_mem_size_cells, &usm);
716 }
717 nid = of_drconf_to_nid_single(&drmem, &aa);
718 fake_numa_create_new_node(
719 ((base + size) >> PAGE_SHIFT),
720 &nid);
721 node_set_online(nid);
722 sz = numa_enforce_memory_limit(base, size);
723 if (sz)
724 memblock_set_node(base, sz,
725 &memblock.memory, nid);
726 } while (--ranges);
727 }
728}
729
730static int __init parse_numa_properties(void)
731{
732 struct device_node *memory;
733 int default_nid = 0;
734 unsigned long i;
735
736 if (numa_enabled == 0) {
737 printk(KERN_WARNING "NUMA disabled by user\n");
738 return -1;
739 }
740
741 min_common_depth = find_min_common_depth();
742
743 if (min_common_depth < 0)
744 return min_common_depth;
745
746 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
747
748 /*
749 * Even though we connect cpus to numa domains later in SMP
750 * init, we need to know the node ids now. This is because
751 * each node to be onlined must have NODE_DATA etc backing it.
752 */
753 for_each_present_cpu(i) {
754 struct device_node *cpu;
755 int nid;
756
757 cpu = of_get_cpu_node(i, NULL);
758 BUG_ON(!cpu);
759 nid = of_node_to_nid_single(cpu);
760 of_node_put(cpu);
761
762 /*
763 * Don't fall back to default_nid yet -- we will plug
764 * cpus into nodes once the memory scan has discovered
765 * the topology.
766 */
767 if (nid < 0)
768 continue;
769 node_set_online(nid);
770 }
771
772 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
773
774 for_each_node_by_type(memory, "memory") {
775 unsigned long start;
776 unsigned long size;
777 int nid;
778 int ranges;
779 const __be32 *memcell_buf;
780 unsigned int len;
781
782 memcell_buf = of_get_property(memory,
783 "linux,usable-memory", &len);
784 if (!memcell_buf || len <= 0)
785 memcell_buf = of_get_property(memory, "reg", &len);
786 if (!memcell_buf || len <= 0)
787 continue;
788
789 /* ranges in cell */
790 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
791new_range:
792 /* these are order-sensitive, and modify the buffer pointer */
793 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
794 size = read_n_cells(n_mem_size_cells, &memcell_buf);
795
796 /*
797 * Assumption: either all memory nodes or none will
798 * have associativity properties. If none, then
799 * everything goes to default_nid.
800 */
801 nid = of_node_to_nid_single(memory);
802 if (nid < 0)
803 nid = default_nid;
804
805 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
806 node_set_online(nid);
807
808 if (!(size = numa_enforce_memory_limit(start, size))) {
809 if (--ranges)
810 goto new_range;
811 else
812 continue;
813 }
814
815 memblock_set_node(start, size, &memblock.memory, nid);
816
817 if (--ranges)
818 goto new_range;
819 }
820
821 /*
822 * Now do the same thing for each MEMBLOCK listed in the
823 * ibm,dynamic-memory property in the
824 * ibm,dynamic-reconfiguration-memory node.
825 */
826 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
827 if (memory)
828 parse_drconf_memory(memory);
829
830 return 0;
831}
832
833static void __init setup_nonnuma(void)
834{
835 unsigned long top_of_ram = memblock_end_of_DRAM();
836 unsigned long total_ram = memblock_phys_mem_size();
837 unsigned long start_pfn, end_pfn;
838 unsigned int nid = 0;
839 struct memblock_region *reg;
840
841 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
842 top_of_ram, total_ram);
843 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
844 (top_of_ram - total_ram) >> 20);
845
846 for_each_memblock(memory, reg) {
847 start_pfn = memblock_region_memory_base_pfn(reg);
848 end_pfn = memblock_region_memory_end_pfn(reg);
849
850 fake_numa_create_new_node(end_pfn, &nid);
851 memblock_set_node(PFN_PHYS(start_pfn),
852 PFN_PHYS(end_pfn - start_pfn),
853 &memblock.memory, nid);
854 node_set_online(nid);
855 }
856}
857
858void __init dump_numa_cpu_topology(void)
859{
860 unsigned int node;
861 unsigned int cpu, count;
862
863 if (min_common_depth == -1 || !numa_enabled)
864 return;
865
866 for_each_online_node(node) {
867 printk(KERN_DEBUG "Node %d CPUs:", node);
868
869 count = 0;
870 /*
871 * If we used a CPU iterator here we would miss printing
872 * the holes in the cpumap.
873 */
874 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
875 if (cpumask_test_cpu(cpu,
876 node_to_cpumask_map[node])) {
877 if (count == 0)
878 printk(" %u", cpu);
879 ++count;
880 } else {
881 if (count > 1)
882 printk("-%u", cpu - 1);
883 count = 0;
884 }
885 }
886
887 if (count > 1)
888 printk("-%u", nr_cpu_ids - 1);
889 printk("\n");
890 }
891}
892
893static void __init dump_numa_memory_topology(void)
894{
895 unsigned int node;
896 unsigned int count;
897
898 if (min_common_depth == -1 || !numa_enabled)
899 return;
900
901 for_each_online_node(node) {
902 unsigned long i;
903
904 printk(KERN_DEBUG "Node %d Memory:", node);
905
906 count = 0;
907
908 for (i = 0; i < memblock_end_of_DRAM();
909 i += (1 << SECTION_SIZE_BITS)) {
910 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
911 if (count == 0)
912 printk(" 0x%lx", i);
913 ++count;
914 } else {
915 if (count > 0)
916 printk("-0x%lx", i);
917 count = 0;
918 }
919 }
920
921 if (count > 0)
922 printk("-0x%lx", i);
923 printk("\n");
924 }
925}
926
927/*
928 * Allocate some memory, satisfying the memblock or bootmem allocator where
929 * required. nid is the preferred node and end is the physical address of
930 * the highest address in the node.
931 *
932 * Returns the virtual address of the memory.
933 */
934static void __init *careful_zallocation(int nid, unsigned long size,
935 unsigned long align,
936 unsigned long end_pfn)
937{
938 void *ret;
939 int new_nid;
940 unsigned long ret_paddr;
941
942 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
943
944 /* retry over all memory */
945 if (!ret_paddr)
946 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
947
948 if (!ret_paddr)
949 panic("numa.c: cannot allocate %lu bytes for node %d",
950 size, nid);
951
952 ret = __va(ret_paddr);
953
954 /*
955 * We initialize the nodes in numeric order: 0, 1, 2...
956 * and hand over control from the MEMBLOCK allocator to the
957 * bootmem allocator. If this function is called for
958 * node 5, then we know that all nodes <5 are using the
959 * bootmem allocator instead of the MEMBLOCK allocator.
960 *
961 * So, check the nid from which this allocation came
962 * and double check to see if we need to use bootmem
963 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
964 * since it would be useless.
965 */
966 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
967 if (new_nid < nid) {
968 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
969 size, align, 0);
970
971 dbg("alloc_bootmem %p %lx\n", ret, size);
972 }
973
974 memset(ret, 0, size);
975 return ret;
976}
977
978static struct notifier_block ppc64_numa_nb = {
979 .notifier_call = cpu_numa_callback,
980 .priority = 1 /* Must run before sched domains notifier. */
981};
982
983static void __init mark_reserved_regions_for_nid(int nid)
984{
985 struct pglist_data *node = NODE_DATA(nid);
986 struct memblock_region *reg;
987
988 for_each_memblock(reserved, reg) {
989 unsigned long physbase = reg->base;
990 unsigned long size = reg->size;
991 unsigned long start_pfn = physbase >> PAGE_SHIFT;
992 unsigned long end_pfn = PFN_UP(physbase + size);
993 struct node_active_region node_ar;
994 unsigned long node_end_pfn = pgdat_end_pfn(node);
995
996 /*
997 * Check to make sure that this memblock.reserved area is
998 * within the bounds of the node that we care about.
999 * Checking the nid of the start and end points is not
1000 * sufficient because the reserved area could span the
1001 * entire node.
1002 */
1003 if (end_pfn <= node->node_start_pfn ||
1004 start_pfn >= node_end_pfn)
1005 continue;
1006
1007 get_node_active_region(start_pfn, &node_ar);
1008 while (start_pfn < end_pfn &&
1009 node_ar.start_pfn < node_ar.end_pfn) {
1010 unsigned long reserve_size = size;
1011 /*
1012 * if reserved region extends past active region
1013 * then trim size to active region
1014 */
1015 if (end_pfn > node_ar.end_pfn)
1016 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1017 - physbase;
1018 /*
1019 * Only worry about *this* node, others may not
1020 * yet have valid NODE_DATA().
1021 */
1022 if (node_ar.nid == nid) {
1023 dbg("reserve_bootmem %lx %lx nid=%d\n",
1024 physbase, reserve_size, node_ar.nid);
1025 reserve_bootmem_node(NODE_DATA(node_ar.nid),
1026 physbase, reserve_size,
1027 BOOTMEM_DEFAULT);
1028 }
1029 /*
1030 * if reserved region is contained in the active region
1031 * then done.
1032 */
1033 if (end_pfn <= node_ar.end_pfn)
1034 break;
1035
1036 /*
1037 * reserved region extends past the active region
1038 * get next active region that contains this
1039 * reserved region
1040 */
1041 start_pfn = node_ar.end_pfn;
1042 physbase = start_pfn << PAGE_SHIFT;
1043 size = size - reserve_size;
1044 get_node_active_region(start_pfn, &node_ar);
1045 }
1046 }
1047}
1048
1049
1050void __init do_init_bootmem(void)
1051{
1052 int nid;
1053
1054 min_low_pfn = 0;
1055 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1056 max_pfn = max_low_pfn;
1057
1058 if (parse_numa_properties())
1059 setup_nonnuma();
1060 else
1061 dump_numa_memory_topology();
1062
1063 for_each_online_node(nid) {
1064 unsigned long start_pfn, end_pfn;
1065 void *bootmem_vaddr;
1066 unsigned long bootmap_pages;
1067
1068 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1069
1070 /*
1071 * Allocate the node structure node local if possible
1072 *
1073 * Be careful moving this around, as it relies on all
1074 * previous nodes' bootmem to be initialized and have
1075 * all reserved areas marked.
1076 */
1077 NODE_DATA(nid) = careful_zallocation(nid,
1078 sizeof(struct pglist_data),
1079 SMP_CACHE_BYTES, end_pfn);
1080
1081 dbg("node %d\n", nid);
1082 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1083
1084 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1085 NODE_DATA(nid)->node_start_pfn = start_pfn;
1086 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1087
1088 if (NODE_DATA(nid)->node_spanned_pages == 0)
1089 continue;
1090
1091 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1092 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1093
1094 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1095 bootmem_vaddr = careful_zallocation(nid,
1096 bootmap_pages << PAGE_SHIFT,
1097 PAGE_SIZE, end_pfn);
1098
1099 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1100
1101 init_bootmem_node(NODE_DATA(nid),
1102 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1103 start_pfn, end_pfn);
1104
1105 free_bootmem_with_active_regions(nid, end_pfn);
1106 /*
1107 * Be very careful about moving this around. Future
1108 * calls to careful_zallocation() depend on this getting
1109 * done correctly.
1110 */
1111 mark_reserved_regions_for_nid(nid);
1112 sparse_memory_present_with_active_regions(nid);
1113 }
1114
1115 init_bootmem_done = 1;
1116
1117 /*
1118 * Now bootmem is initialised we can create the node to cpumask
1119 * lookup tables and setup the cpu callback to populate them.
1120 */
1121 setup_node_to_cpumask_map();
1122
1123 reset_numa_cpu_lookup_table();
1124 register_cpu_notifier(&ppc64_numa_nb);
1125 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1126 (void *)(unsigned long)boot_cpuid);
1127}
1128
1129void __init paging_init(void)
1130{
1131 unsigned long max_zone_pfns[MAX_NR_ZONES];
1132 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1133 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1134 free_area_init_nodes(max_zone_pfns);
1135}
1136
1137static int __init early_numa(char *p)
1138{
1139 if (!p)
1140 return 0;
1141
1142 if (strstr(p, "off"))
1143 numa_enabled = 0;
1144
1145 if (strstr(p, "debug"))
1146 numa_debug = 1;
1147
1148 p = strstr(p, "fake=");
1149 if (p)
1150 cmdline = p + strlen("fake=");
1151
1152 return 0;
1153}
1154early_param("numa", early_numa);
1155
1156#ifdef CONFIG_MEMORY_HOTPLUG
1157/*
1158 * Find the node associated with a hot added memory section for
1159 * memory represented in the device tree by the property
1160 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1161 */
1162static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1163 unsigned long scn_addr)
1164{
1165 const __be32 *dm;
1166 unsigned int drconf_cell_cnt, rc;
1167 unsigned long lmb_size;
1168 struct assoc_arrays aa;
1169 int nid = -1;
1170
1171 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1172 if (!drconf_cell_cnt)
1173 return -1;
1174
1175 lmb_size = of_get_lmb_size(memory);
1176 if (!lmb_size)
1177 return -1;
1178
1179 rc = of_get_assoc_arrays(memory, &aa);
1180 if (rc)
1181 return -1;
1182
1183 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1184 struct of_drconf_cell drmem;
1185
1186 read_drconf_cell(&drmem, &dm);
1187
1188 /* skip this block if it is reserved or not assigned to
1189 * this partition */
1190 if ((drmem.flags & DRCONF_MEM_RESERVED)
1191 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1192 continue;
1193
1194 if ((scn_addr < drmem.base_addr)
1195 || (scn_addr >= (drmem.base_addr + lmb_size)))
1196 continue;
1197
1198 nid = of_drconf_to_nid_single(&drmem, &aa);
1199 break;
1200 }
1201
1202 return nid;
1203}
1204
1205/*
1206 * Find the node associated with a hot added memory section for memory
1207 * represented in the device tree as a node (i.e. memory@XXXX) for
1208 * each memblock.
1209 */
1210static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1211{
1212 struct device_node *memory;
1213 int nid = -1;
1214
1215 for_each_node_by_type(memory, "memory") {
1216 unsigned long start, size;
1217 int ranges;
1218 const __be32 *memcell_buf;
1219 unsigned int len;
1220
1221 memcell_buf = of_get_property(memory, "reg", &len);
1222 if (!memcell_buf || len <= 0)
1223 continue;
1224
1225 /* ranges in cell */
1226 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1227
1228 while (ranges--) {
1229 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1230 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1231
1232 if ((scn_addr < start) || (scn_addr >= (start + size)))
1233 continue;
1234
1235 nid = of_node_to_nid_single(memory);
1236 break;
1237 }
1238
1239 if (nid >= 0)
1240 break;
1241 }
1242
1243 of_node_put(memory);
1244
1245 return nid;
1246}
1247
1248/*
1249 * Find the node associated with a hot added memory section. Section
1250 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1251 * sections are fully contained within a single MEMBLOCK.
1252 */
1253int hot_add_scn_to_nid(unsigned long scn_addr)
1254{
1255 struct device_node *memory = NULL;
1256 int nid, found = 0;
1257
1258 if (!numa_enabled || (min_common_depth < 0))
1259 return first_online_node;
1260
1261 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1262 if (memory) {
1263 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1264 of_node_put(memory);
1265 } else {
1266 nid = hot_add_node_scn_to_nid(scn_addr);
1267 }
1268
1269 if (nid < 0 || !node_online(nid))
1270 nid = first_online_node;
1271
1272 if (NODE_DATA(nid)->node_spanned_pages)
1273 return nid;
1274
1275 for_each_online_node(nid) {
1276 if (NODE_DATA(nid)->node_spanned_pages) {
1277 found = 1;
1278 break;
1279 }
1280 }
1281
1282 BUG_ON(!found);
1283 return nid;
1284}
1285
1286static u64 hot_add_drconf_memory_max(void)
1287{
1288 struct device_node *memory = NULL;
1289 unsigned int drconf_cell_cnt = 0;
1290 u64 lmb_size = 0;
1291 const __be32 *dm = NULL;
1292
1293 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1294 if (memory) {
1295 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1296 lmb_size = of_get_lmb_size(memory);
1297 of_node_put(memory);
1298 }
1299 return lmb_size * drconf_cell_cnt;
1300}
1301
1302/*
1303 * memory_hotplug_max - return max address of memory that may be added
1304 *
1305 * This is currently only used on systems that support drconfig memory
1306 * hotplug.
1307 */
1308u64 memory_hotplug_max(void)
1309{
1310 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1311}
1312#endif /* CONFIG_MEMORY_HOTPLUG */
1313
1314/* Virtual Processor Home Node (VPHN) support */
1315#ifdef CONFIG_PPC_SPLPAR
1316struct topology_update_data {
1317 struct topology_update_data *next;
1318 unsigned int cpu;
1319 int old_nid;
1320 int new_nid;
1321};
1322
1323static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1324static cpumask_t cpu_associativity_changes_mask;
1325static int vphn_enabled;
1326static int prrn_enabled;
1327static void reset_topology_timer(void);
1328
1329/*
1330 * Store the current values of the associativity change counters in the
1331 * hypervisor.
1332 */
1333static void setup_cpu_associativity_change_counters(void)
1334{
1335 int cpu;
1336
1337 /* The VPHN feature supports a maximum of 8 reference points */
1338 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1339
1340 for_each_possible_cpu(cpu) {
1341 int i;
1342 u8 *counts = vphn_cpu_change_counts[cpu];
1343 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1344
1345 for (i = 0; i < distance_ref_points_depth; i++)
1346 counts[i] = hypervisor_counts[i];
1347 }
1348}
1349
1350/*
1351 * The hypervisor maintains a set of 8 associativity change counters in
1352 * the VPA of each cpu that correspond to the associativity levels in the
1353 * ibm,associativity-reference-points property. When an associativity
1354 * level changes, the corresponding counter is incremented.
1355 *
1356 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1357 * node associativity levels have changed.
1358 *
1359 * Returns the number of cpus with unhandled associativity changes.
1360 */
1361static int update_cpu_associativity_changes_mask(void)
1362{
1363 int cpu;
1364 cpumask_t *changes = &cpu_associativity_changes_mask;
1365
1366 for_each_possible_cpu(cpu) {
1367 int i, changed = 0;
1368 u8 *counts = vphn_cpu_change_counts[cpu];
1369 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1370
1371 for (i = 0; i < distance_ref_points_depth; i++) {
1372 if (hypervisor_counts[i] != counts[i]) {
1373 counts[i] = hypervisor_counts[i];
1374 changed = 1;
1375 }
1376 }
1377 if (changed) {
1378 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1379 cpu = cpu_last_thread_sibling(cpu);
1380 }
1381 }
1382
1383 return cpumask_weight(changes);
1384}
1385
1386/*
1387 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1388 * the complete property we have to add the length in the first cell.
1389 */
1390#define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1391
1392/*
1393 * Convert the associativity domain numbers returned from the hypervisor
1394 * to the sequence they would appear in the ibm,associativity property.
1395 */
1396static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1397{
1398 int i, nr_assoc_doms = 0;
1399 const __be16 *field = (const __be16 *) packed;
1400
1401#define VPHN_FIELD_UNUSED (0xffff)
1402#define VPHN_FIELD_MSB (0x8000)
1403#define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1404
1405 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1406 if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1407 /* All significant fields processed, and remaining
1408 * fields contain the reserved value of all 1's.
1409 * Just store them.
1410 */
1411 unpacked[i] = *((__be32 *)field);
1412 field += 2;
1413 } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1414 /* Data is in the lower 15 bits of this field */
1415 unpacked[i] = cpu_to_be32(
1416 be16_to_cpup(field) & VPHN_FIELD_MASK);
1417 field++;
1418 nr_assoc_doms++;
1419 } else {
1420 /* Data is in the lower 15 bits of this field
1421 * concatenated with the next 16 bit field
1422 */
1423 unpacked[i] = *((__be32 *)field);
1424 field += 2;
1425 nr_assoc_doms++;
1426 }
1427 }
1428
1429 /* The first cell contains the length of the property */
1430 unpacked[0] = cpu_to_be32(nr_assoc_doms);
1431
1432 return nr_assoc_doms;
1433}
1434
1435/*
1436 * Retrieve the new associativity information for a virtual processor's
1437 * home node.
1438 */
1439static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1440{
1441 long rc;
1442 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1443 u64 flags = 1;
1444 int hwcpu = get_hard_smp_processor_id(cpu);
1445
1446 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1447 vphn_unpack_associativity(retbuf, associativity);
1448
1449 return rc;
1450}
1451
1452static long vphn_get_associativity(unsigned long cpu,
1453 __be32 *associativity)
1454{
1455 long rc;
1456
1457 rc = hcall_vphn(cpu, associativity);
1458
1459 switch (rc) {
1460 case H_FUNCTION:
1461 printk(KERN_INFO
1462 "VPHN is not supported. Disabling polling...\n");
1463 stop_topology_update();
1464 break;
1465 case H_HARDWARE:
1466 printk(KERN_ERR
1467 "hcall_vphn() experienced a hardware fault "
1468 "preventing VPHN. Disabling polling...\n");
1469 stop_topology_update();
1470 }
1471
1472 return rc;
1473}
1474
1475/*
1476 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1477 * characteristics change. This function doesn't perform any locking and is
1478 * only safe to call from stop_machine().
1479 */
1480static int update_cpu_topology(void *data)
1481{
1482 struct topology_update_data *update;
1483 unsigned long cpu;
1484
1485 if (!data)
1486 return -EINVAL;
1487
1488 cpu = smp_processor_id();
1489
1490 for (update = data; update; update = update->next) {
1491 if (cpu != update->cpu)
1492 continue;
1493
1494 unmap_cpu_from_node(update->cpu);
1495 map_cpu_to_node(update->cpu, update->new_nid);
1496 vdso_getcpu_init();
1497 }
1498
1499 return 0;
1500}
1501
1502static int update_lookup_table(void *data)
1503{
1504 struct topology_update_data *update;
1505
1506 if (!data)
1507 return -EINVAL;
1508
1509 /*
1510 * Upon topology update, the numa-cpu lookup table needs to be updated
1511 * for all threads in the core, including offline CPUs, to ensure that
1512 * future hotplug operations respect the cpu-to-node associativity
1513 * properly.
1514 */
1515 for (update = data; update; update = update->next) {
1516 int nid, base, j;
1517
1518 nid = update->new_nid;
1519 base = cpu_first_thread_sibling(update->cpu);
1520
1521 for (j = 0; j < threads_per_core; j++) {
1522 update_numa_cpu_lookup_table(base + j, nid);
1523 }
1524 }
1525
1526 return 0;
1527}
1528
1529/*
1530 * Update the node maps and sysfs entries for each cpu whose home node
1531 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1532 */
1533int arch_update_cpu_topology(void)
1534{
1535 unsigned int cpu, sibling, changed = 0;
1536 struct topology_update_data *updates, *ud;
1537 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1538 cpumask_t updated_cpus;
1539 struct device *dev;
1540 int weight, new_nid, i = 0;
1541
1542 weight = cpumask_weight(&cpu_associativity_changes_mask);
1543 if (!weight)
1544 return 0;
1545
1546 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1547 if (!updates)
1548 return 0;
1549
1550 cpumask_clear(&updated_cpus);
1551
1552 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1553 /*
1554 * If siblings aren't flagged for changes, updates list
1555 * will be too short. Skip on this update and set for next
1556 * update.
1557 */
1558 if (!cpumask_subset(cpu_sibling_mask(cpu),
1559 &cpu_associativity_changes_mask)) {
1560 pr_info("Sibling bits not set for associativity "
1561 "change, cpu%d\n", cpu);
1562 cpumask_or(&cpu_associativity_changes_mask,
1563 &cpu_associativity_changes_mask,
1564 cpu_sibling_mask(cpu));
1565 cpu = cpu_last_thread_sibling(cpu);
1566 continue;
1567 }
1568
1569 /* Use associativity from first thread for all siblings */
1570 vphn_get_associativity(cpu, associativity);
1571 new_nid = associativity_to_nid(associativity);
1572 if (new_nid < 0 || !node_online(new_nid))
1573 new_nid = first_online_node;
1574
1575 if (new_nid == numa_cpu_lookup_table[cpu]) {
1576 cpumask_andnot(&cpu_associativity_changes_mask,
1577 &cpu_associativity_changes_mask,
1578 cpu_sibling_mask(cpu));
1579 cpu = cpu_last_thread_sibling(cpu);
1580 continue;
1581 }
1582
1583 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1584 ud = &updates[i++];
1585 ud->cpu = sibling;
1586 ud->new_nid = new_nid;
1587 ud->old_nid = numa_cpu_lookup_table[sibling];
1588 cpumask_set_cpu(sibling, &updated_cpus);
1589 if (i < weight)
1590 ud->next = &updates[i];
1591 }
1592 cpu = cpu_last_thread_sibling(cpu);
1593 }
1594
1595 /*
1596 * In cases where we have nothing to update (because the updates list
1597 * is too short or because the new topology is same as the old one),
1598 * skip invoking update_cpu_topology() via stop-machine(). This is
1599 * necessary (and not just a fast-path optimization) since stop-machine
1600 * can end up electing a random CPU to run update_cpu_topology(), and
1601 * thus trick us into setting up incorrect cpu-node mappings (since
1602 * 'updates' is kzalloc()'ed).
1603 *
1604 * And for the similar reason, we will skip all the following updating.
1605 */
1606 if (!cpumask_weight(&updated_cpus))
1607 goto out;
1608
1609 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1610
1611 /*
1612 * Update the numa-cpu lookup table with the new mappings, even for
1613 * offline CPUs. It is best to perform this update from the stop-
1614 * machine context.
1615 */
1616 stop_machine(update_lookup_table, &updates[0],
1617 cpumask_of(raw_smp_processor_id()));
1618
1619 for (ud = &updates[0]; ud; ud = ud->next) {
1620 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1621 register_cpu_under_node(ud->cpu, ud->new_nid);
1622
1623 dev = get_cpu_device(ud->cpu);
1624 if (dev)
1625 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1626 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1627 changed = 1;
1628 }
1629
1630out:
1631 kfree(updates);
1632 return changed;
1633}
1634
1635static void topology_work_fn(struct work_struct *work)
1636{
1637 rebuild_sched_domains();
1638}
1639static DECLARE_WORK(topology_work, topology_work_fn);
1640
1641static void topology_schedule_update(void)
1642{
1643 schedule_work(&topology_work);
1644}
1645
1646static void topology_timer_fn(unsigned long ignored)
1647{
1648 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1649 topology_schedule_update();
1650 else if (vphn_enabled) {
1651 if (update_cpu_associativity_changes_mask() > 0)
1652 topology_schedule_update();
1653 reset_topology_timer();
1654 }
1655}
1656static struct timer_list topology_timer =
1657 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1658
1659static void reset_topology_timer(void)
1660{
1661 topology_timer.data = 0;
1662 topology_timer.expires = jiffies + 60 * HZ;
1663 mod_timer(&topology_timer, topology_timer.expires);
1664}
1665
1666#ifdef CONFIG_SMP
1667
1668static void stage_topology_update(int core_id)
1669{
1670 cpumask_or(&cpu_associativity_changes_mask,
1671 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1672 reset_topology_timer();
1673}
1674
1675static int dt_update_callback(struct notifier_block *nb,
1676 unsigned long action, void *data)
1677{
1678 struct of_prop_reconfig *update;
1679 int rc = NOTIFY_DONE;
1680
1681 switch (action) {
1682 case OF_RECONFIG_UPDATE_PROPERTY:
1683 update = (struct of_prop_reconfig *)data;
1684 if (!of_prop_cmp(update->dn->type, "cpu") &&
1685 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1686 u32 core_id;
1687 of_property_read_u32(update->dn, "reg", &core_id);
1688 stage_topology_update(core_id);
1689 rc = NOTIFY_OK;
1690 }
1691 break;
1692 }
1693
1694 return rc;
1695}
1696
1697static struct notifier_block dt_update_nb = {
1698 .notifier_call = dt_update_callback,
1699};
1700
1701#endif
1702
1703/*
1704 * Start polling for associativity changes.
1705 */
1706int start_topology_update(void)
1707{
1708 int rc = 0;
1709
1710 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1711 if (!prrn_enabled) {
1712 prrn_enabled = 1;
1713 vphn_enabled = 0;
1714#ifdef CONFIG_SMP
1715 rc = of_reconfig_notifier_register(&dt_update_nb);
1716#endif
1717 }
1718 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1719 lppaca_shared_proc(get_lppaca())) {
1720 if (!vphn_enabled) {
1721 prrn_enabled = 0;
1722 vphn_enabled = 1;
1723 setup_cpu_associativity_change_counters();
1724 init_timer_deferrable(&topology_timer);
1725 reset_topology_timer();
1726 }
1727 }
1728
1729 return rc;
1730}
1731
1732/*
1733 * Disable polling for VPHN associativity changes.
1734 */
1735int stop_topology_update(void)
1736{
1737 int rc = 0;
1738
1739 if (prrn_enabled) {
1740 prrn_enabled = 0;
1741#ifdef CONFIG_SMP
1742 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1743#endif
1744 } else if (vphn_enabled) {
1745 vphn_enabled = 0;
1746 rc = del_timer_sync(&topology_timer);
1747 }
1748
1749 return rc;
1750}
1751
1752int prrn_is_enabled(void)
1753{
1754 return prrn_enabled;
1755}
1756
1757static int topology_read(struct seq_file *file, void *v)
1758{
1759 if (vphn_enabled || prrn_enabled)
1760 seq_puts(file, "on\n");
1761 else
1762 seq_puts(file, "off\n");
1763
1764 return 0;
1765}
1766
1767static int topology_open(struct inode *inode, struct file *file)
1768{
1769 return single_open(file, topology_read, NULL);
1770}
1771
1772static ssize_t topology_write(struct file *file, const char __user *buf,
1773 size_t count, loff_t *off)
1774{
1775 char kbuf[4]; /* "on" or "off" plus null. */
1776 int read_len;
1777
1778 read_len = count < 3 ? count : 3;
1779 if (copy_from_user(kbuf, buf, read_len))
1780 return -EINVAL;
1781
1782 kbuf[read_len] = '\0';
1783
1784 if (!strncmp(kbuf, "on", 2))
1785 start_topology_update();
1786 else if (!strncmp(kbuf, "off", 3))
1787 stop_topology_update();
1788 else
1789 return -EINVAL;
1790
1791 return count;
1792}
1793
1794static const struct file_operations topology_ops = {
1795 .read = seq_read,
1796 .write = topology_write,
1797 .open = topology_open,
1798 .release = single_release
1799};
1800
1801static int topology_update_init(void)
1802{
1803 start_topology_update();
1804 proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops);
1805
1806 return 0;
1807}
1808device_initcall(topology_update_init);
1809#endif /* CONFIG_PPC_SPLPAR */