xref: /openbmc/linux/arch/powerpc/mm/numa.c (revision e23feb16)
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/firmware.h>
35 #include <asm/paca.h>
36 #include <asm/hvcall.h>
37 #include <asm/setup.h>
38 #include <asm/vdso.h>
39 
40 static int numa_enabled = 1;
41 
42 static char *cmdline __initdata;
43 
44 static int numa_debug;
45 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
46 
47 int numa_cpu_lookup_table[NR_CPUS];
48 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
49 struct pglist_data *node_data[MAX_NUMNODES];
50 
51 EXPORT_SYMBOL(numa_cpu_lookup_table);
52 EXPORT_SYMBOL(node_to_cpumask_map);
53 EXPORT_SYMBOL(node_data);
54 
55 static int min_common_depth;
56 static int n_mem_addr_cells, n_mem_size_cells;
57 static int form1_affinity;
58 
59 #define MAX_DISTANCE_REF_POINTS 4
60 static int distance_ref_points_depth;
61 static const __be32 *distance_ref_points;
62 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
63 
64 /*
65  * Allocate node_to_cpumask_map based on number of available nodes
66  * Requires node_possible_map to be valid.
67  *
68  * Note: cpumask_of_node() is not valid until after this is done.
69  */
70 static void __init setup_node_to_cpumask_map(void)
71 {
72 	unsigned int node;
73 
74 	/* setup nr_node_ids if not done yet */
75 	if (nr_node_ids == MAX_NUMNODES)
76 		setup_nr_node_ids();
77 
78 	/* allocate the map */
79 	for (node = 0; node < nr_node_ids; node++)
80 		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
81 
82 	/* cpumask_of_node() will now work */
83 	dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
84 }
85 
86 static int __init fake_numa_create_new_node(unsigned long end_pfn,
87 						unsigned int *nid)
88 {
89 	unsigned long long mem;
90 	char *p = cmdline;
91 	static unsigned int fake_nid;
92 	static unsigned long long curr_boundary;
93 
94 	/*
95 	 * Modify node id, iff we started creating NUMA nodes
96 	 * We want to continue from where we left of the last time
97 	 */
98 	if (fake_nid)
99 		*nid = fake_nid;
100 	/*
101 	 * In case there are no more arguments to parse, the
102 	 * node_id should be the same as the last fake node id
103 	 * (we've handled this above).
104 	 */
105 	if (!p)
106 		return 0;
107 
108 	mem = memparse(p, &p);
109 	if (!mem)
110 		return 0;
111 
112 	if (mem < curr_boundary)
113 		return 0;
114 
115 	curr_boundary = mem;
116 
117 	if ((end_pfn << PAGE_SHIFT) > mem) {
118 		/*
119 		 * Skip commas and spaces
120 		 */
121 		while (*p == ',' || *p == ' ' || *p == '\t')
122 			p++;
123 
124 		cmdline = p;
125 		fake_nid++;
126 		*nid = fake_nid;
127 		dbg("created new fake_node with id %d\n", fake_nid);
128 		return 1;
129 	}
130 	return 0;
131 }
132 
133 /*
134  * get_node_active_region - Return active region containing pfn
135  * Active range returned is empty if none found.
136  * @pfn: The page to return the region for
137  * @node_ar: Returned set to the active region containing @pfn
138  */
139 static void __init get_node_active_region(unsigned long pfn,
140 					  struct node_active_region *node_ar)
141 {
142 	unsigned long start_pfn, end_pfn;
143 	int i, nid;
144 
145 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
146 		if (pfn >= start_pfn && pfn < end_pfn) {
147 			node_ar->nid = nid;
148 			node_ar->start_pfn = start_pfn;
149 			node_ar->end_pfn = end_pfn;
150 			break;
151 		}
152 	}
153 }
154 
155 static void map_cpu_to_node(int cpu, int node)
156 {
157 	numa_cpu_lookup_table[cpu] = node;
158 
159 	dbg("adding cpu %d to node %d\n", cpu, node);
160 
161 	if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
162 		cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
163 }
164 
165 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
166 static void unmap_cpu_from_node(unsigned long cpu)
167 {
168 	int node = numa_cpu_lookup_table[cpu];
169 
170 	dbg("removing cpu %lu from node %d\n", cpu, node);
171 
172 	if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
173 		cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
174 	} else {
175 		printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
176 		       cpu, node);
177 	}
178 }
179 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
180 
181 /* must hold reference to node during call */
182 static const __be32 *of_get_associativity(struct device_node *dev)
183 {
184 	return of_get_property(dev, "ibm,associativity", NULL);
185 }
186 
187 /*
188  * Returns the property linux,drconf-usable-memory if
189  * it exists (the property exists only in kexec/kdump kernels,
190  * added by kexec-tools)
191  */
192 static const __be32 *of_get_usable_memory(struct device_node *memory)
193 {
194 	const __be32 *prop;
195 	u32 len;
196 	prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
197 	if (!prop || len < sizeof(unsigned int))
198 		return 0;
199 	return prop;
200 }
201 
202 int __node_distance(int a, int b)
203 {
204 	int i;
205 	int distance = LOCAL_DISTANCE;
206 
207 	if (!form1_affinity)
208 		return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
209 
210 	for (i = 0; i < distance_ref_points_depth; i++) {
211 		if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
212 			break;
213 
214 		/* Double the distance for each NUMA level */
215 		distance *= 2;
216 	}
217 
218 	return distance;
219 }
220 
221 static void initialize_distance_lookup_table(int nid,
222 		const __be32 *associativity)
223 {
224 	int i;
225 
226 	if (!form1_affinity)
227 		return;
228 
229 	for (i = 0; i < distance_ref_points_depth; i++) {
230 		const __be32 *entry;
231 
232 		entry = &associativity[be32_to_cpu(distance_ref_points[i])];
233 		distance_lookup_table[nid][i] = of_read_number(entry, 1);
234 	}
235 }
236 
237 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
238  * info is found.
239  */
240 static int associativity_to_nid(const __be32 *associativity)
241 {
242 	int nid = -1;
243 
244 	if (min_common_depth == -1)
245 		goto out;
246 
247 	if (of_read_number(associativity, 1) >= min_common_depth)
248 		nid = of_read_number(&associativity[min_common_depth], 1);
249 
250 	/* POWER4 LPAR uses 0xffff as invalid node */
251 	if (nid == 0xffff || nid >= MAX_NUMNODES)
252 		nid = -1;
253 
254 	if (nid > 0 &&
255 	    of_read_number(associativity, 1) >= distance_ref_points_depth)
256 		initialize_distance_lookup_table(nid, associativity);
257 
258 out:
259 	return nid;
260 }
261 
262 /* Returns the nid associated with the given device tree node,
263  * or -1 if not found.
264  */
265 static int of_node_to_nid_single(struct device_node *device)
266 {
267 	int nid = -1;
268 	const __be32 *tmp;
269 
270 	tmp = of_get_associativity(device);
271 	if (tmp)
272 		nid = associativity_to_nid(tmp);
273 	return nid;
274 }
275 
276 /* Walk the device tree upwards, looking for an associativity id */
277 int of_node_to_nid(struct device_node *device)
278 {
279 	struct device_node *tmp;
280 	int nid = -1;
281 
282 	of_node_get(device);
283 	while (device) {
284 		nid = of_node_to_nid_single(device);
285 		if (nid != -1)
286 			break;
287 
288 	        tmp = device;
289 		device = of_get_parent(tmp);
290 		of_node_put(tmp);
291 	}
292 	of_node_put(device);
293 
294 	return nid;
295 }
296 EXPORT_SYMBOL_GPL(of_node_to_nid);
297 
298 static int __init find_min_common_depth(void)
299 {
300 	int depth;
301 	struct device_node *root;
302 
303 	if (firmware_has_feature(FW_FEATURE_OPAL))
304 		root = of_find_node_by_path("/ibm,opal");
305 	else
306 		root = of_find_node_by_path("/rtas");
307 	if (!root)
308 		root = of_find_node_by_path("/");
309 
310 	/*
311 	 * This property is a set of 32-bit integers, each representing
312 	 * an index into the ibm,associativity nodes.
313 	 *
314 	 * With form 0 affinity the first integer is for an SMP configuration
315 	 * (should be all 0's) and the second is for a normal NUMA
316 	 * configuration. We have only one level of NUMA.
317 	 *
318 	 * With form 1 affinity the first integer is the most significant
319 	 * NUMA boundary and the following are progressively less significant
320 	 * boundaries. There can be more than one level of NUMA.
321 	 */
322 	distance_ref_points = of_get_property(root,
323 					"ibm,associativity-reference-points",
324 					&distance_ref_points_depth);
325 
326 	if (!distance_ref_points) {
327 		dbg("NUMA: ibm,associativity-reference-points not found.\n");
328 		goto err;
329 	}
330 
331 	distance_ref_points_depth /= sizeof(int);
332 
333 	if (firmware_has_feature(FW_FEATURE_OPAL) ||
334 	    firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
335 		dbg("Using form 1 affinity\n");
336 		form1_affinity = 1;
337 	}
338 
339 	if (form1_affinity) {
340 		depth = of_read_number(distance_ref_points, 1);
341 	} else {
342 		if (distance_ref_points_depth < 2) {
343 			printk(KERN_WARNING "NUMA: "
344 				"short ibm,associativity-reference-points\n");
345 			goto err;
346 		}
347 
348 		depth = of_read_number(&distance_ref_points[1], 1);
349 	}
350 
351 	/*
352 	 * Warn and cap if the hardware supports more than
353 	 * MAX_DISTANCE_REF_POINTS domains.
354 	 */
355 	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
356 		printk(KERN_WARNING "NUMA: distance array capped at "
357 			"%d entries\n", MAX_DISTANCE_REF_POINTS);
358 		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
359 	}
360 
361 	of_node_put(root);
362 	return depth;
363 
364 err:
365 	of_node_put(root);
366 	return -1;
367 }
368 
369 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
370 {
371 	struct device_node *memory = NULL;
372 
373 	memory = of_find_node_by_type(memory, "memory");
374 	if (!memory)
375 		panic("numa.c: No memory nodes found!");
376 
377 	*n_addr_cells = of_n_addr_cells(memory);
378 	*n_size_cells = of_n_size_cells(memory);
379 	of_node_put(memory);
380 }
381 
382 static unsigned long read_n_cells(int n, const __be32 **buf)
383 {
384 	unsigned long result = 0;
385 
386 	while (n--) {
387 		result = (result << 32) | of_read_number(*buf, 1);
388 		(*buf)++;
389 	}
390 	return result;
391 }
392 
393 /*
394  * Read the next memblock list entry from the ibm,dynamic-memory property
395  * and return the information in the provided of_drconf_cell structure.
396  */
397 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
398 {
399 	const __be32 *cp;
400 
401 	drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
402 
403 	cp = *cellp;
404 	drmem->drc_index = of_read_number(cp, 1);
405 	drmem->reserved = of_read_number(&cp[1], 1);
406 	drmem->aa_index = of_read_number(&cp[2], 1);
407 	drmem->flags = of_read_number(&cp[3], 1);
408 
409 	*cellp = cp + 4;
410 }
411 
412 /*
413  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
414  *
415  * The layout of the ibm,dynamic-memory property is a number N of memblock
416  * list entries followed by N memblock list entries.  Each memblock list entry
417  * contains information as laid out in the of_drconf_cell struct above.
418  */
419 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
420 {
421 	const __be32 *prop;
422 	u32 len, entries;
423 
424 	prop = of_get_property(memory, "ibm,dynamic-memory", &len);
425 	if (!prop || len < sizeof(unsigned int))
426 		return 0;
427 
428 	entries = of_read_number(prop++, 1);
429 
430 	/* Now that we know the number of entries, revalidate the size
431 	 * of the property read in to ensure we have everything
432 	 */
433 	if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
434 		return 0;
435 
436 	*dm = prop;
437 	return entries;
438 }
439 
440 /*
441  * Retrieve and validate the ibm,lmb-size property for drconf memory
442  * from the device tree.
443  */
444 static u64 of_get_lmb_size(struct device_node *memory)
445 {
446 	const __be32 *prop;
447 	u32 len;
448 
449 	prop = of_get_property(memory, "ibm,lmb-size", &len);
450 	if (!prop || len < sizeof(unsigned int))
451 		return 0;
452 
453 	return read_n_cells(n_mem_size_cells, &prop);
454 }
455 
456 struct assoc_arrays {
457 	u32	n_arrays;
458 	u32	array_sz;
459 	const __be32 *arrays;
460 };
461 
462 /*
463  * Retrieve and validate the list of associativity arrays for drconf
464  * memory from the ibm,associativity-lookup-arrays property of the
465  * device tree..
466  *
467  * The layout of the ibm,associativity-lookup-arrays property is a number N
468  * indicating the number of associativity arrays, followed by a number M
469  * indicating the size of each associativity array, followed by a list
470  * of N associativity arrays.
471  */
472 static int of_get_assoc_arrays(struct device_node *memory,
473 			       struct assoc_arrays *aa)
474 {
475 	const __be32 *prop;
476 	u32 len;
477 
478 	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
479 	if (!prop || len < 2 * sizeof(unsigned int))
480 		return -1;
481 
482 	aa->n_arrays = of_read_number(prop++, 1);
483 	aa->array_sz = of_read_number(prop++, 1);
484 
485 	/* Now that we know the number of arrays and size of each array,
486 	 * revalidate the size of the property read in.
487 	 */
488 	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
489 		return -1;
490 
491 	aa->arrays = prop;
492 	return 0;
493 }
494 
495 /*
496  * This is like of_node_to_nid_single() for memory represented in the
497  * ibm,dynamic-reconfiguration-memory node.
498  */
499 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
500 				   struct assoc_arrays *aa)
501 {
502 	int default_nid = 0;
503 	int nid = default_nid;
504 	int index;
505 
506 	if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
507 	    !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
508 	    drmem->aa_index < aa->n_arrays) {
509 		index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
510 		nid = of_read_number(&aa->arrays[index], 1);
511 
512 		if (nid == 0xffff || nid >= MAX_NUMNODES)
513 			nid = default_nid;
514 	}
515 
516 	return nid;
517 }
518 
519 /*
520  * Figure out to which domain a cpu belongs and stick it there.
521  * Return the id of the domain used.
522  */
523 static int numa_setup_cpu(unsigned long lcpu)
524 {
525 	int nid = 0;
526 	struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
527 
528 	if (!cpu) {
529 		WARN_ON(1);
530 		goto out;
531 	}
532 
533 	nid = of_node_to_nid_single(cpu);
534 
535 	if (nid < 0 || !node_online(nid))
536 		nid = first_online_node;
537 out:
538 	map_cpu_to_node(lcpu, nid);
539 
540 	of_node_put(cpu);
541 
542 	return nid;
543 }
544 
545 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
546 			     void *hcpu)
547 {
548 	unsigned long lcpu = (unsigned long)hcpu;
549 	int ret = NOTIFY_DONE;
550 
551 	switch (action) {
552 	case CPU_UP_PREPARE:
553 	case CPU_UP_PREPARE_FROZEN:
554 		numa_setup_cpu(lcpu);
555 		ret = NOTIFY_OK;
556 		break;
557 #ifdef CONFIG_HOTPLUG_CPU
558 	case CPU_DEAD:
559 	case CPU_DEAD_FROZEN:
560 	case CPU_UP_CANCELED:
561 	case CPU_UP_CANCELED_FROZEN:
562 		unmap_cpu_from_node(lcpu);
563 		break;
564 		ret = NOTIFY_OK;
565 #endif
566 	}
567 	return ret;
568 }
569 
570 /*
571  * Check and possibly modify a memory region to enforce the memory limit.
572  *
573  * Returns the size the region should have to enforce the memory limit.
574  * This will either be the original value of size, a truncated value,
575  * or zero. If the returned value of size is 0 the region should be
576  * discarded as it lies wholly above the memory limit.
577  */
578 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
579 						      unsigned long size)
580 {
581 	/*
582 	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
583 	 * we've already adjusted it for the limit and it takes care of
584 	 * having memory holes below the limit.  Also, in the case of
585 	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
586 	 */
587 
588 	if (start + size <= memblock_end_of_DRAM())
589 		return size;
590 
591 	if (start >= memblock_end_of_DRAM())
592 		return 0;
593 
594 	return memblock_end_of_DRAM() - start;
595 }
596 
597 /*
598  * Reads the counter for a given entry in
599  * linux,drconf-usable-memory property
600  */
601 static inline int __init read_usm_ranges(const __be32 **usm)
602 {
603 	/*
604 	 * For each lmb in ibm,dynamic-memory a corresponding
605 	 * entry in linux,drconf-usable-memory property contains
606 	 * a counter followed by that many (base, size) duple.
607 	 * read the counter from linux,drconf-usable-memory
608 	 */
609 	return read_n_cells(n_mem_size_cells, usm);
610 }
611 
612 /*
613  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
614  * node.  This assumes n_mem_{addr,size}_cells have been set.
615  */
616 static void __init parse_drconf_memory(struct device_node *memory)
617 {
618 	const __be32 *uninitialized_var(dm), *usm;
619 	unsigned int n, rc, ranges, is_kexec_kdump = 0;
620 	unsigned long lmb_size, base, size, sz;
621 	int nid;
622 	struct assoc_arrays aa = { .arrays = NULL };
623 
624 	n = of_get_drconf_memory(memory, &dm);
625 	if (!n)
626 		return;
627 
628 	lmb_size = of_get_lmb_size(memory);
629 	if (!lmb_size)
630 		return;
631 
632 	rc = of_get_assoc_arrays(memory, &aa);
633 	if (rc)
634 		return;
635 
636 	/* check if this is a kexec/kdump kernel */
637 	usm = of_get_usable_memory(memory);
638 	if (usm != NULL)
639 		is_kexec_kdump = 1;
640 
641 	for (; n != 0; --n) {
642 		struct of_drconf_cell drmem;
643 
644 		read_drconf_cell(&drmem, &dm);
645 
646 		/* skip this block if the reserved bit is set in flags (0x80)
647 		   or if the block is not assigned to this partition (0x8) */
648 		if ((drmem.flags & DRCONF_MEM_RESERVED)
649 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
650 			continue;
651 
652 		base = drmem.base_addr;
653 		size = lmb_size;
654 		ranges = 1;
655 
656 		if (is_kexec_kdump) {
657 			ranges = read_usm_ranges(&usm);
658 			if (!ranges) /* there are no (base, size) duple */
659 				continue;
660 		}
661 		do {
662 			if (is_kexec_kdump) {
663 				base = read_n_cells(n_mem_addr_cells, &usm);
664 				size = read_n_cells(n_mem_size_cells, &usm);
665 			}
666 			nid = of_drconf_to_nid_single(&drmem, &aa);
667 			fake_numa_create_new_node(
668 				((base + size) >> PAGE_SHIFT),
669 					   &nid);
670 			node_set_online(nid);
671 			sz = numa_enforce_memory_limit(base, size);
672 			if (sz)
673 				memblock_set_node(base, sz, nid);
674 		} while (--ranges);
675 	}
676 }
677 
678 static int __init parse_numa_properties(void)
679 {
680 	struct device_node *memory;
681 	int default_nid = 0;
682 	unsigned long i;
683 
684 	if (numa_enabled == 0) {
685 		printk(KERN_WARNING "NUMA disabled by user\n");
686 		return -1;
687 	}
688 
689 	min_common_depth = find_min_common_depth();
690 
691 	if (min_common_depth < 0)
692 		return min_common_depth;
693 
694 	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
695 
696 	/*
697 	 * Even though we connect cpus to numa domains later in SMP
698 	 * init, we need to know the node ids now. This is because
699 	 * each node to be onlined must have NODE_DATA etc backing it.
700 	 */
701 	for_each_present_cpu(i) {
702 		struct device_node *cpu;
703 		int nid;
704 
705 		cpu = of_get_cpu_node(i, NULL);
706 		BUG_ON(!cpu);
707 		nid = of_node_to_nid_single(cpu);
708 		of_node_put(cpu);
709 
710 		/*
711 		 * Don't fall back to default_nid yet -- we will plug
712 		 * cpus into nodes once the memory scan has discovered
713 		 * the topology.
714 		 */
715 		if (nid < 0)
716 			continue;
717 		node_set_online(nid);
718 	}
719 
720 	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
721 
722 	for_each_node_by_type(memory, "memory") {
723 		unsigned long start;
724 		unsigned long size;
725 		int nid;
726 		int ranges;
727 		const __be32 *memcell_buf;
728 		unsigned int len;
729 
730 		memcell_buf = of_get_property(memory,
731 			"linux,usable-memory", &len);
732 		if (!memcell_buf || len <= 0)
733 			memcell_buf = of_get_property(memory, "reg", &len);
734 		if (!memcell_buf || len <= 0)
735 			continue;
736 
737 		/* ranges in cell */
738 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
739 new_range:
740 		/* these are order-sensitive, and modify the buffer pointer */
741 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
742 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
743 
744 		/*
745 		 * Assumption: either all memory nodes or none will
746 		 * have associativity properties.  If none, then
747 		 * everything goes to default_nid.
748 		 */
749 		nid = of_node_to_nid_single(memory);
750 		if (nid < 0)
751 			nid = default_nid;
752 
753 		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
754 		node_set_online(nid);
755 
756 		if (!(size = numa_enforce_memory_limit(start, size))) {
757 			if (--ranges)
758 				goto new_range;
759 			else
760 				continue;
761 		}
762 
763 		memblock_set_node(start, size, nid);
764 
765 		if (--ranges)
766 			goto new_range;
767 	}
768 
769 	/*
770 	 * Now do the same thing for each MEMBLOCK listed in the
771 	 * ibm,dynamic-memory property in the
772 	 * ibm,dynamic-reconfiguration-memory node.
773 	 */
774 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
775 	if (memory)
776 		parse_drconf_memory(memory);
777 
778 	return 0;
779 }
780 
781 static void __init setup_nonnuma(void)
782 {
783 	unsigned long top_of_ram = memblock_end_of_DRAM();
784 	unsigned long total_ram = memblock_phys_mem_size();
785 	unsigned long start_pfn, end_pfn;
786 	unsigned int nid = 0;
787 	struct memblock_region *reg;
788 
789 	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
790 	       top_of_ram, total_ram);
791 	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
792 	       (top_of_ram - total_ram) >> 20);
793 
794 	for_each_memblock(memory, reg) {
795 		start_pfn = memblock_region_memory_base_pfn(reg);
796 		end_pfn = memblock_region_memory_end_pfn(reg);
797 
798 		fake_numa_create_new_node(end_pfn, &nid);
799 		memblock_set_node(PFN_PHYS(start_pfn),
800 				  PFN_PHYS(end_pfn - start_pfn), nid);
801 		node_set_online(nid);
802 	}
803 }
804 
805 void __init dump_numa_cpu_topology(void)
806 {
807 	unsigned int node;
808 	unsigned int cpu, count;
809 
810 	if (min_common_depth == -1 || !numa_enabled)
811 		return;
812 
813 	for_each_online_node(node) {
814 		printk(KERN_DEBUG "Node %d CPUs:", node);
815 
816 		count = 0;
817 		/*
818 		 * If we used a CPU iterator here we would miss printing
819 		 * the holes in the cpumap.
820 		 */
821 		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
822 			if (cpumask_test_cpu(cpu,
823 					node_to_cpumask_map[node])) {
824 				if (count == 0)
825 					printk(" %u", cpu);
826 				++count;
827 			} else {
828 				if (count > 1)
829 					printk("-%u", cpu - 1);
830 				count = 0;
831 			}
832 		}
833 
834 		if (count > 1)
835 			printk("-%u", nr_cpu_ids - 1);
836 		printk("\n");
837 	}
838 }
839 
840 static void __init dump_numa_memory_topology(void)
841 {
842 	unsigned int node;
843 	unsigned int count;
844 
845 	if (min_common_depth == -1 || !numa_enabled)
846 		return;
847 
848 	for_each_online_node(node) {
849 		unsigned long i;
850 
851 		printk(KERN_DEBUG "Node %d Memory:", node);
852 
853 		count = 0;
854 
855 		for (i = 0; i < memblock_end_of_DRAM();
856 		     i += (1 << SECTION_SIZE_BITS)) {
857 			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
858 				if (count == 0)
859 					printk(" 0x%lx", i);
860 				++count;
861 			} else {
862 				if (count > 0)
863 					printk("-0x%lx", i);
864 				count = 0;
865 			}
866 		}
867 
868 		if (count > 0)
869 			printk("-0x%lx", i);
870 		printk("\n");
871 	}
872 }
873 
874 /*
875  * Allocate some memory, satisfying the memblock or bootmem allocator where
876  * required. nid is the preferred node and end is the physical address of
877  * the highest address in the node.
878  *
879  * Returns the virtual address of the memory.
880  */
881 static void __init *careful_zallocation(int nid, unsigned long size,
882 				       unsigned long align,
883 				       unsigned long end_pfn)
884 {
885 	void *ret;
886 	int new_nid;
887 	unsigned long ret_paddr;
888 
889 	ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
890 
891 	/* retry over all memory */
892 	if (!ret_paddr)
893 		ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
894 
895 	if (!ret_paddr)
896 		panic("numa.c: cannot allocate %lu bytes for node %d",
897 		      size, nid);
898 
899 	ret = __va(ret_paddr);
900 
901 	/*
902 	 * We initialize the nodes in numeric order: 0, 1, 2...
903 	 * and hand over control from the MEMBLOCK allocator to the
904 	 * bootmem allocator.  If this function is called for
905 	 * node 5, then we know that all nodes <5 are using the
906 	 * bootmem allocator instead of the MEMBLOCK allocator.
907 	 *
908 	 * So, check the nid from which this allocation came
909 	 * and double check to see if we need to use bootmem
910 	 * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
911 	 * since it would be useless.
912 	 */
913 	new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
914 	if (new_nid < nid) {
915 		ret = __alloc_bootmem_node(NODE_DATA(new_nid),
916 				size, align, 0);
917 
918 		dbg("alloc_bootmem %p %lx\n", ret, size);
919 	}
920 
921 	memset(ret, 0, size);
922 	return ret;
923 }
924 
925 static struct notifier_block ppc64_numa_nb = {
926 	.notifier_call = cpu_numa_callback,
927 	.priority = 1 /* Must run before sched domains notifier. */
928 };
929 
930 static void __init mark_reserved_regions_for_nid(int nid)
931 {
932 	struct pglist_data *node = NODE_DATA(nid);
933 	struct memblock_region *reg;
934 
935 	for_each_memblock(reserved, reg) {
936 		unsigned long physbase = reg->base;
937 		unsigned long size = reg->size;
938 		unsigned long start_pfn = physbase >> PAGE_SHIFT;
939 		unsigned long end_pfn = PFN_UP(physbase + size);
940 		struct node_active_region node_ar;
941 		unsigned long node_end_pfn = node->node_start_pfn +
942 					     node->node_spanned_pages;
943 
944 		/*
945 		 * Check to make sure that this memblock.reserved area is
946 		 * within the bounds of the node that we care about.
947 		 * Checking the nid of the start and end points is not
948 		 * sufficient because the reserved area could span the
949 		 * entire node.
950 		 */
951 		if (end_pfn <= node->node_start_pfn ||
952 		    start_pfn >= node_end_pfn)
953 			continue;
954 
955 		get_node_active_region(start_pfn, &node_ar);
956 		while (start_pfn < end_pfn &&
957 			node_ar.start_pfn < node_ar.end_pfn) {
958 			unsigned long reserve_size = size;
959 			/*
960 			 * if reserved region extends past active region
961 			 * then trim size to active region
962 			 */
963 			if (end_pfn > node_ar.end_pfn)
964 				reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
965 					- physbase;
966 			/*
967 			 * Only worry about *this* node, others may not
968 			 * yet have valid NODE_DATA().
969 			 */
970 			if (node_ar.nid == nid) {
971 				dbg("reserve_bootmem %lx %lx nid=%d\n",
972 					physbase, reserve_size, node_ar.nid);
973 				reserve_bootmem_node(NODE_DATA(node_ar.nid),
974 						physbase, reserve_size,
975 						BOOTMEM_DEFAULT);
976 			}
977 			/*
978 			 * if reserved region is contained in the active region
979 			 * then done.
980 			 */
981 			if (end_pfn <= node_ar.end_pfn)
982 				break;
983 
984 			/*
985 			 * reserved region extends past the active region
986 			 *   get next active region that contains this
987 			 *   reserved region
988 			 */
989 			start_pfn = node_ar.end_pfn;
990 			physbase = start_pfn << PAGE_SHIFT;
991 			size = size - reserve_size;
992 			get_node_active_region(start_pfn, &node_ar);
993 		}
994 	}
995 }
996 
997 
998 void __init do_init_bootmem(void)
999 {
1000 	int nid;
1001 
1002 	min_low_pfn = 0;
1003 	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1004 	max_pfn = max_low_pfn;
1005 
1006 	if (parse_numa_properties())
1007 		setup_nonnuma();
1008 	else
1009 		dump_numa_memory_topology();
1010 
1011 	for_each_online_node(nid) {
1012 		unsigned long start_pfn, end_pfn;
1013 		void *bootmem_vaddr;
1014 		unsigned long bootmap_pages;
1015 
1016 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1017 
1018 		/*
1019 		 * Allocate the node structure node local if possible
1020 		 *
1021 		 * Be careful moving this around, as it relies on all
1022 		 * previous nodes' bootmem to be initialized and have
1023 		 * all reserved areas marked.
1024 		 */
1025 		NODE_DATA(nid) = careful_zallocation(nid,
1026 					sizeof(struct pglist_data),
1027 					SMP_CACHE_BYTES, end_pfn);
1028 
1029   		dbg("node %d\n", nid);
1030 		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1031 
1032 		NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1033 		NODE_DATA(nid)->node_start_pfn = start_pfn;
1034 		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1035 
1036 		if (NODE_DATA(nid)->node_spanned_pages == 0)
1037   			continue;
1038 
1039   		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1040   		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1041 
1042 		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1043 		bootmem_vaddr = careful_zallocation(nid,
1044 					bootmap_pages << PAGE_SHIFT,
1045 					PAGE_SIZE, end_pfn);
1046 
1047 		dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1048 
1049 		init_bootmem_node(NODE_DATA(nid),
1050 				  __pa(bootmem_vaddr) >> PAGE_SHIFT,
1051 				  start_pfn, end_pfn);
1052 
1053 		free_bootmem_with_active_regions(nid, end_pfn);
1054 		/*
1055 		 * Be very careful about moving this around.  Future
1056 		 * calls to careful_zallocation() depend on this getting
1057 		 * done correctly.
1058 		 */
1059 		mark_reserved_regions_for_nid(nid);
1060 		sparse_memory_present_with_active_regions(nid);
1061 	}
1062 
1063 	init_bootmem_done = 1;
1064 
1065 	/*
1066 	 * Now bootmem is initialised we can create the node to cpumask
1067 	 * lookup tables and setup the cpu callback to populate them.
1068 	 */
1069 	setup_node_to_cpumask_map();
1070 
1071 	register_cpu_notifier(&ppc64_numa_nb);
1072 	cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1073 			  (void *)(unsigned long)boot_cpuid);
1074 }
1075 
1076 void __init paging_init(void)
1077 {
1078 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1079 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1080 	max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1081 	free_area_init_nodes(max_zone_pfns);
1082 }
1083 
1084 static int __init early_numa(char *p)
1085 {
1086 	if (!p)
1087 		return 0;
1088 
1089 	if (strstr(p, "off"))
1090 		numa_enabled = 0;
1091 
1092 	if (strstr(p, "debug"))
1093 		numa_debug = 1;
1094 
1095 	p = strstr(p, "fake=");
1096 	if (p)
1097 		cmdline = p + strlen("fake=");
1098 
1099 	return 0;
1100 }
1101 early_param("numa", early_numa);
1102 
1103 #ifdef CONFIG_MEMORY_HOTPLUG
1104 /*
1105  * Find the node associated with a hot added memory section for
1106  * memory represented in the device tree by the property
1107  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1108  */
1109 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1110 				     unsigned long scn_addr)
1111 {
1112 	const __be32 *dm;
1113 	unsigned int drconf_cell_cnt, rc;
1114 	unsigned long lmb_size;
1115 	struct assoc_arrays aa;
1116 	int nid = -1;
1117 
1118 	drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1119 	if (!drconf_cell_cnt)
1120 		return -1;
1121 
1122 	lmb_size = of_get_lmb_size(memory);
1123 	if (!lmb_size)
1124 		return -1;
1125 
1126 	rc = of_get_assoc_arrays(memory, &aa);
1127 	if (rc)
1128 		return -1;
1129 
1130 	for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1131 		struct of_drconf_cell drmem;
1132 
1133 		read_drconf_cell(&drmem, &dm);
1134 
1135 		/* skip this block if it is reserved or not assigned to
1136 		 * this partition */
1137 		if ((drmem.flags & DRCONF_MEM_RESERVED)
1138 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1139 			continue;
1140 
1141 		if ((scn_addr < drmem.base_addr)
1142 		    || (scn_addr >= (drmem.base_addr + lmb_size)))
1143 			continue;
1144 
1145 		nid = of_drconf_to_nid_single(&drmem, &aa);
1146 		break;
1147 	}
1148 
1149 	return nid;
1150 }
1151 
1152 /*
1153  * Find the node associated with a hot added memory section for memory
1154  * represented in the device tree as a node (i.e. memory@XXXX) for
1155  * each memblock.
1156  */
1157 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1158 {
1159 	struct device_node *memory;
1160 	int nid = -1;
1161 
1162 	for_each_node_by_type(memory, "memory") {
1163 		unsigned long start, size;
1164 		int ranges;
1165 		const __be32 *memcell_buf;
1166 		unsigned int len;
1167 
1168 		memcell_buf = of_get_property(memory, "reg", &len);
1169 		if (!memcell_buf || len <= 0)
1170 			continue;
1171 
1172 		/* ranges in cell */
1173 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1174 
1175 		while (ranges--) {
1176 			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1177 			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1178 
1179 			if ((scn_addr < start) || (scn_addr >= (start + size)))
1180 				continue;
1181 
1182 			nid = of_node_to_nid_single(memory);
1183 			break;
1184 		}
1185 
1186 		if (nid >= 0)
1187 			break;
1188 	}
1189 
1190 	of_node_put(memory);
1191 
1192 	return nid;
1193 }
1194 
1195 /*
1196  * Find the node associated with a hot added memory section.  Section
1197  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1198  * sections are fully contained within a single MEMBLOCK.
1199  */
1200 int hot_add_scn_to_nid(unsigned long scn_addr)
1201 {
1202 	struct device_node *memory = NULL;
1203 	int nid, found = 0;
1204 
1205 	if (!numa_enabled || (min_common_depth < 0))
1206 		return first_online_node;
1207 
1208 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1209 	if (memory) {
1210 		nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1211 		of_node_put(memory);
1212 	} else {
1213 		nid = hot_add_node_scn_to_nid(scn_addr);
1214 	}
1215 
1216 	if (nid < 0 || !node_online(nid))
1217 		nid = first_online_node;
1218 
1219 	if (NODE_DATA(nid)->node_spanned_pages)
1220 		return nid;
1221 
1222 	for_each_online_node(nid) {
1223 		if (NODE_DATA(nid)->node_spanned_pages) {
1224 			found = 1;
1225 			break;
1226 		}
1227 	}
1228 
1229 	BUG_ON(!found);
1230 	return nid;
1231 }
1232 
1233 static u64 hot_add_drconf_memory_max(void)
1234 {
1235         struct device_node *memory = NULL;
1236         unsigned int drconf_cell_cnt = 0;
1237         u64 lmb_size = 0;
1238 	const __be32 *dm = 0;
1239 
1240         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1241         if (memory) {
1242                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1243                 lmb_size = of_get_lmb_size(memory);
1244                 of_node_put(memory);
1245         }
1246         return lmb_size * drconf_cell_cnt;
1247 }
1248 
1249 /*
1250  * memory_hotplug_max - return max address of memory that may be added
1251  *
1252  * This is currently only used on systems that support drconfig memory
1253  * hotplug.
1254  */
1255 u64 memory_hotplug_max(void)
1256 {
1257         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1258 }
1259 #endif /* CONFIG_MEMORY_HOTPLUG */
1260 
1261 /* Virtual Processor Home Node (VPHN) support */
1262 #ifdef CONFIG_PPC_SPLPAR
1263 struct topology_update_data {
1264 	struct topology_update_data *next;
1265 	unsigned int cpu;
1266 	int old_nid;
1267 	int new_nid;
1268 };
1269 
1270 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1271 static cpumask_t cpu_associativity_changes_mask;
1272 static int vphn_enabled;
1273 static int prrn_enabled;
1274 static void reset_topology_timer(void);
1275 
1276 /*
1277  * Store the current values of the associativity change counters in the
1278  * hypervisor.
1279  */
1280 static void setup_cpu_associativity_change_counters(void)
1281 {
1282 	int cpu;
1283 
1284 	/* The VPHN feature supports a maximum of 8 reference points */
1285 	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1286 
1287 	for_each_possible_cpu(cpu) {
1288 		int i;
1289 		u8 *counts = vphn_cpu_change_counts[cpu];
1290 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1291 
1292 		for (i = 0; i < distance_ref_points_depth; i++)
1293 			counts[i] = hypervisor_counts[i];
1294 	}
1295 }
1296 
1297 /*
1298  * The hypervisor maintains a set of 8 associativity change counters in
1299  * the VPA of each cpu that correspond to the associativity levels in the
1300  * ibm,associativity-reference-points property. When an associativity
1301  * level changes, the corresponding counter is incremented.
1302  *
1303  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1304  * node associativity levels have changed.
1305  *
1306  * Returns the number of cpus with unhandled associativity changes.
1307  */
1308 static int update_cpu_associativity_changes_mask(void)
1309 {
1310 	int cpu;
1311 	cpumask_t *changes = &cpu_associativity_changes_mask;
1312 
1313 	for_each_possible_cpu(cpu) {
1314 		int i, changed = 0;
1315 		u8 *counts = vphn_cpu_change_counts[cpu];
1316 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1317 
1318 		for (i = 0; i < distance_ref_points_depth; i++) {
1319 			if (hypervisor_counts[i] != counts[i]) {
1320 				counts[i] = hypervisor_counts[i];
1321 				changed = 1;
1322 			}
1323 		}
1324 		if (changed) {
1325 			cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1326 			cpu = cpu_last_thread_sibling(cpu);
1327 		}
1328 	}
1329 
1330 	return cpumask_weight(changes);
1331 }
1332 
1333 /*
1334  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1335  * the complete property we have to add the length in the first cell.
1336  */
1337 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1338 
1339 /*
1340  * Convert the associativity domain numbers returned from the hypervisor
1341  * to the sequence they would appear in the ibm,associativity property.
1342  */
1343 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1344 {
1345 	int i, nr_assoc_doms = 0;
1346 	const __be16 *field = (const __be16 *) packed;
1347 
1348 #define VPHN_FIELD_UNUSED	(0xffff)
1349 #define VPHN_FIELD_MSB		(0x8000)
1350 #define VPHN_FIELD_MASK		(~VPHN_FIELD_MSB)
1351 
1352 	for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1353 		if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1354 			/* All significant fields processed, and remaining
1355 			 * fields contain the reserved value of all 1's.
1356 			 * Just store them.
1357 			 */
1358 			unpacked[i] = *((__be32 *)field);
1359 			field += 2;
1360 		} else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1361 			/* Data is in the lower 15 bits of this field */
1362 			unpacked[i] = cpu_to_be32(
1363 				be16_to_cpup(field) & VPHN_FIELD_MASK);
1364 			field++;
1365 			nr_assoc_doms++;
1366 		} else {
1367 			/* Data is in the lower 15 bits of this field
1368 			 * concatenated with the next 16 bit field
1369 			 */
1370 			unpacked[i] = *((__be32 *)field);
1371 			field += 2;
1372 			nr_assoc_doms++;
1373 		}
1374 	}
1375 
1376 	/* The first cell contains the length of the property */
1377 	unpacked[0] = cpu_to_be32(nr_assoc_doms);
1378 
1379 	return nr_assoc_doms;
1380 }
1381 
1382 /*
1383  * Retrieve the new associativity information for a virtual processor's
1384  * home node.
1385  */
1386 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1387 {
1388 	long rc;
1389 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1390 	u64 flags = 1;
1391 	int hwcpu = get_hard_smp_processor_id(cpu);
1392 
1393 	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1394 	vphn_unpack_associativity(retbuf, associativity);
1395 
1396 	return rc;
1397 }
1398 
1399 static long vphn_get_associativity(unsigned long cpu,
1400 					__be32 *associativity)
1401 {
1402 	long rc;
1403 
1404 	rc = hcall_vphn(cpu, associativity);
1405 
1406 	switch (rc) {
1407 	case H_FUNCTION:
1408 		printk(KERN_INFO
1409 			"VPHN is not supported. Disabling polling...\n");
1410 		stop_topology_update();
1411 		break;
1412 	case H_HARDWARE:
1413 		printk(KERN_ERR
1414 			"hcall_vphn() experienced a hardware fault "
1415 			"preventing VPHN. Disabling polling...\n");
1416 		stop_topology_update();
1417 	}
1418 
1419 	return rc;
1420 }
1421 
1422 /*
1423  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1424  * characteristics change. This function doesn't perform any locking and is
1425  * only safe to call from stop_machine().
1426  */
1427 static int update_cpu_topology(void *data)
1428 {
1429 	struct topology_update_data *update;
1430 	unsigned long cpu;
1431 
1432 	if (!data)
1433 		return -EINVAL;
1434 
1435 	cpu = smp_processor_id();
1436 
1437 	for (update = data; update; update = update->next) {
1438 		if (cpu != update->cpu)
1439 			continue;
1440 
1441 		unmap_cpu_from_node(update->cpu);
1442 		map_cpu_to_node(update->cpu, update->new_nid);
1443 		vdso_getcpu_init();
1444 	}
1445 
1446 	return 0;
1447 }
1448 
1449 /*
1450  * Update the node maps and sysfs entries for each cpu whose home node
1451  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1452  */
1453 int arch_update_cpu_topology(void)
1454 {
1455 	unsigned int cpu, sibling, changed = 0;
1456 	struct topology_update_data *updates, *ud;
1457 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1458 	cpumask_t updated_cpus;
1459 	struct device *dev;
1460 	int weight, new_nid, i = 0;
1461 
1462 	weight = cpumask_weight(&cpu_associativity_changes_mask);
1463 	if (!weight)
1464 		return 0;
1465 
1466 	updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1467 	if (!updates)
1468 		return 0;
1469 
1470 	cpumask_clear(&updated_cpus);
1471 
1472 	for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1473 		/*
1474 		 * If siblings aren't flagged for changes, updates list
1475 		 * will be too short. Skip on this update and set for next
1476 		 * update.
1477 		 */
1478 		if (!cpumask_subset(cpu_sibling_mask(cpu),
1479 					&cpu_associativity_changes_mask)) {
1480 			pr_info("Sibling bits not set for associativity "
1481 					"change, cpu%d\n", cpu);
1482 			cpumask_or(&cpu_associativity_changes_mask,
1483 					&cpu_associativity_changes_mask,
1484 					cpu_sibling_mask(cpu));
1485 			cpu = cpu_last_thread_sibling(cpu);
1486 			continue;
1487 		}
1488 
1489 		/* Use associativity from first thread for all siblings */
1490 		vphn_get_associativity(cpu, associativity);
1491 		new_nid = associativity_to_nid(associativity);
1492 		if (new_nid < 0 || !node_online(new_nid))
1493 			new_nid = first_online_node;
1494 
1495 		if (new_nid == numa_cpu_lookup_table[cpu]) {
1496 			cpumask_andnot(&cpu_associativity_changes_mask,
1497 					&cpu_associativity_changes_mask,
1498 					cpu_sibling_mask(cpu));
1499 			cpu = cpu_last_thread_sibling(cpu);
1500 			continue;
1501 		}
1502 
1503 		for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1504 			ud = &updates[i++];
1505 			ud->cpu = sibling;
1506 			ud->new_nid = new_nid;
1507 			ud->old_nid = numa_cpu_lookup_table[sibling];
1508 			cpumask_set_cpu(sibling, &updated_cpus);
1509 			if (i < weight)
1510 				ud->next = &updates[i];
1511 		}
1512 		cpu = cpu_last_thread_sibling(cpu);
1513 	}
1514 
1515 	stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1516 
1517 	for (ud = &updates[0]; ud; ud = ud->next) {
1518 		unregister_cpu_under_node(ud->cpu, ud->old_nid);
1519 		register_cpu_under_node(ud->cpu, ud->new_nid);
1520 
1521 		dev = get_cpu_device(ud->cpu);
1522 		if (dev)
1523 			kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1524 		cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1525 		changed = 1;
1526 	}
1527 
1528 	kfree(updates);
1529 	return changed;
1530 }
1531 
1532 static void topology_work_fn(struct work_struct *work)
1533 {
1534 	rebuild_sched_domains();
1535 }
1536 static DECLARE_WORK(topology_work, topology_work_fn);
1537 
1538 void topology_schedule_update(void)
1539 {
1540 	schedule_work(&topology_work);
1541 }
1542 
1543 static void topology_timer_fn(unsigned long ignored)
1544 {
1545 	if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1546 		topology_schedule_update();
1547 	else if (vphn_enabled) {
1548 		if (update_cpu_associativity_changes_mask() > 0)
1549 			topology_schedule_update();
1550 		reset_topology_timer();
1551 	}
1552 }
1553 static struct timer_list topology_timer =
1554 	TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1555 
1556 static void reset_topology_timer(void)
1557 {
1558 	topology_timer.data = 0;
1559 	topology_timer.expires = jiffies + 60 * HZ;
1560 	mod_timer(&topology_timer, topology_timer.expires);
1561 }
1562 
1563 #ifdef CONFIG_SMP
1564 
1565 static void stage_topology_update(int core_id)
1566 {
1567 	cpumask_or(&cpu_associativity_changes_mask,
1568 		&cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1569 	reset_topology_timer();
1570 }
1571 
1572 static int dt_update_callback(struct notifier_block *nb,
1573 				unsigned long action, void *data)
1574 {
1575 	struct of_prop_reconfig *update;
1576 	int rc = NOTIFY_DONE;
1577 
1578 	switch (action) {
1579 	case OF_RECONFIG_UPDATE_PROPERTY:
1580 		update = (struct of_prop_reconfig *)data;
1581 		if (!of_prop_cmp(update->dn->type, "cpu") &&
1582 		    !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1583 			u32 core_id;
1584 			of_property_read_u32(update->dn, "reg", &core_id);
1585 			stage_topology_update(core_id);
1586 			rc = NOTIFY_OK;
1587 		}
1588 		break;
1589 	}
1590 
1591 	return rc;
1592 }
1593 
1594 static struct notifier_block dt_update_nb = {
1595 	.notifier_call = dt_update_callback,
1596 };
1597 
1598 #endif
1599 
1600 /*
1601  * Start polling for associativity changes.
1602  */
1603 int start_topology_update(void)
1604 {
1605 	int rc = 0;
1606 
1607 	if (firmware_has_feature(FW_FEATURE_PRRN)) {
1608 		if (!prrn_enabled) {
1609 			prrn_enabled = 1;
1610 			vphn_enabled = 0;
1611 #ifdef CONFIG_SMP
1612 			rc = of_reconfig_notifier_register(&dt_update_nb);
1613 #endif
1614 		}
1615 	} else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1616 		   lppaca_shared_proc(get_lppaca())) {
1617 		if (!vphn_enabled) {
1618 			prrn_enabled = 0;
1619 			vphn_enabled = 1;
1620 			setup_cpu_associativity_change_counters();
1621 			init_timer_deferrable(&topology_timer);
1622 			reset_topology_timer();
1623 		}
1624 	}
1625 
1626 	return rc;
1627 }
1628 
1629 /*
1630  * Disable polling for VPHN associativity changes.
1631  */
1632 int stop_topology_update(void)
1633 {
1634 	int rc = 0;
1635 
1636 	if (prrn_enabled) {
1637 		prrn_enabled = 0;
1638 #ifdef CONFIG_SMP
1639 		rc = of_reconfig_notifier_unregister(&dt_update_nb);
1640 #endif
1641 	} else if (vphn_enabled) {
1642 		vphn_enabled = 0;
1643 		rc = del_timer_sync(&topology_timer);
1644 	}
1645 
1646 	return rc;
1647 }
1648 
1649 int prrn_is_enabled(void)
1650 {
1651 	return prrn_enabled;
1652 }
1653 
1654 static int topology_read(struct seq_file *file, void *v)
1655 {
1656 	if (vphn_enabled || prrn_enabled)
1657 		seq_puts(file, "on\n");
1658 	else
1659 		seq_puts(file, "off\n");
1660 
1661 	return 0;
1662 }
1663 
1664 static int topology_open(struct inode *inode, struct file *file)
1665 {
1666 	return single_open(file, topology_read, NULL);
1667 }
1668 
1669 static ssize_t topology_write(struct file *file, const char __user *buf,
1670 			      size_t count, loff_t *off)
1671 {
1672 	char kbuf[4]; /* "on" or "off" plus null. */
1673 	int read_len;
1674 
1675 	read_len = count < 3 ? count : 3;
1676 	if (copy_from_user(kbuf, buf, read_len))
1677 		return -EINVAL;
1678 
1679 	kbuf[read_len] = '\0';
1680 
1681 	if (!strncmp(kbuf, "on", 2))
1682 		start_topology_update();
1683 	else if (!strncmp(kbuf, "off", 3))
1684 		stop_topology_update();
1685 	else
1686 		return -EINVAL;
1687 
1688 	return count;
1689 }
1690 
1691 static const struct file_operations topology_ops = {
1692 	.read = seq_read,
1693 	.write = topology_write,
1694 	.open = topology_open,
1695 	.release = single_release
1696 };
1697 
1698 static int topology_update_init(void)
1699 {
1700 	start_topology_update();
1701 	proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1702 
1703 	return 0;
1704 }
1705 device_initcall(topology_update_init);
1706 #endif /* CONFIG_PPC_SPLPAR */
1707