xref: /openbmc/linux/arch/powerpc/mm/numa.c (revision 1c2f87c2)
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 
42 static int numa_enabled = 1;
43 
44 static char *cmdline __initdata;
45 
46 static int numa_debug;
47 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
48 
49 int numa_cpu_lookup_table[NR_CPUS];
50 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
51 struct pglist_data *node_data[MAX_NUMNODES];
52 
53 EXPORT_SYMBOL(numa_cpu_lookup_table);
54 EXPORT_SYMBOL(node_to_cpumask_map);
55 EXPORT_SYMBOL(node_data);
56 
57 static int min_common_depth;
58 static int n_mem_addr_cells, n_mem_size_cells;
59 static int form1_affinity;
60 
61 #define MAX_DISTANCE_REF_POINTS 4
62 static int distance_ref_points_depth;
63 static const __be32 *distance_ref_points;
64 static 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  */
72 static 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 
88 static 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  */
141 static 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 
157 static 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 
165 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
166 {
167 	numa_cpu_lookup_table[cpu] = node;
168 }
169 
170 static 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)
181 static 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 */
197 static 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  */
207 static 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 
217 int __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 }
235 
236 static void initialize_distance_lookup_table(int nid,
237 		const __be32 *associativity)
238 {
239 	int i;
240 
241 	if (!form1_affinity)
242 		return;
243 
244 	for (i = 0; i < distance_ref_points_depth; i++) {
245 		const __be32 *entry;
246 
247 		entry = &associativity[be32_to_cpu(distance_ref_points[i])];
248 		distance_lookup_table[nid][i] = of_read_number(entry, 1);
249 	}
250 }
251 
252 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
253  * info is found.
254  */
255 static int associativity_to_nid(const __be32 *associativity)
256 {
257 	int nid = -1;
258 
259 	if (min_common_depth == -1)
260 		goto out;
261 
262 	if (of_read_number(associativity, 1) >= min_common_depth)
263 		nid = of_read_number(&associativity[min_common_depth], 1);
264 
265 	/* POWER4 LPAR uses 0xffff as invalid node */
266 	if (nid == 0xffff || nid >= MAX_NUMNODES)
267 		nid = -1;
268 
269 	if (nid > 0 &&
270 	    of_read_number(associativity, 1) >= distance_ref_points_depth)
271 		initialize_distance_lookup_table(nid, associativity);
272 
273 out:
274 	return nid;
275 }
276 
277 /* Returns the nid associated with the given device tree node,
278  * or -1 if not found.
279  */
280 static int of_node_to_nid_single(struct device_node *device)
281 {
282 	int nid = -1;
283 	const __be32 *tmp;
284 
285 	tmp = of_get_associativity(device);
286 	if (tmp)
287 		nid = associativity_to_nid(tmp);
288 	return nid;
289 }
290 
291 /* Walk the device tree upwards, looking for an associativity id */
292 int of_node_to_nid(struct device_node *device)
293 {
294 	struct device_node *tmp;
295 	int nid = -1;
296 
297 	of_node_get(device);
298 	while (device) {
299 		nid = of_node_to_nid_single(device);
300 		if (nid != -1)
301 			break;
302 
303 	        tmp = device;
304 		device = of_get_parent(tmp);
305 		of_node_put(tmp);
306 	}
307 	of_node_put(device);
308 
309 	return nid;
310 }
311 EXPORT_SYMBOL_GPL(of_node_to_nid);
312 
313 static int __init find_min_common_depth(void)
314 {
315 	int depth;
316 	struct device_node *root;
317 
318 	if (firmware_has_feature(FW_FEATURE_OPAL))
319 		root = of_find_node_by_path("/ibm,opal");
320 	else
321 		root = of_find_node_by_path("/rtas");
322 	if (!root)
323 		root = of_find_node_by_path("/");
324 
325 	/*
326 	 * This property is a set of 32-bit integers, each representing
327 	 * an index into the ibm,associativity nodes.
328 	 *
329 	 * With form 0 affinity the first integer is for an SMP configuration
330 	 * (should be all 0's) and the second is for a normal NUMA
331 	 * configuration. We have only one level of NUMA.
332 	 *
333 	 * With form 1 affinity the first integer is the most significant
334 	 * NUMA boundary and the following are progressively less significant
335 	 * boundaries. There can be more than one level of NUMA.
336 	 */
337 	distance_ref_points = of_get_property(root,
338 					"ibm,associativity-reference-points",
339 					&distance_ref_points_depth);
340 
341 	if (!distance_ref_points) {
342 		dbg("NUMA: ibm,associativity-reference-points not found.\n");
343 		goto err;
344 	}
345 
346 	distance_ref_points_depth /= sizeof(int);
347 
348 	if (firmware_has_feature(FW_FEATURE_OPAL) ||
349 	    firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
350 		dbg("Using form 1 affinity\n");
351 		form1_affinity = 1;
352 	}
353 
354 	if (form1_affinity) {
355 		depth = of_read_number(distance_ref_points, 1);
356 	} else {
357 		if (distance_ref_points_depth < 2) {
358 			printk(KERN_WARNING "NUMA: "
359 				"short ibm,associativity-reference-points\n");
360 			goto err;
361 		}
362 
363 		depth = of_read_number(&distance_ref_points[1], 1);
364 	}
365 
366 	/*
367 	 * Warn and cap if the hardware supports more than
368 	 * MAX_DISTANCE_REF_POINTS domains.
369 	 */
370 	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
371 		printk(KERN_WARNING "NUMA: distance array capped at "
372 			"%d entries\n", MAX_DISTANCE_REF_POINTS);
373 		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
374 	}
375 
376 	of_node_put(root);
377 	return depth;
378 
379 err:
380 	of_node_put(root);
381 	return -1;
382 }
383 
384 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
385 {
386 	struct device_node *memory = NULL;
387 
388 	memory = of_find_node_by_type(memory, "memory");
389 	if (!memory)
390 		panic("numa.c: No memory nodes found!");
391 
392 	*n_addr_cells = of_n_addr_cells(memory);
393 	*n_size_cells = of_n_size_cells(memory);
394 	of_node_put(memory);
395 }
396 
397 static unsigned long read_n_cells(int n, const __be32 **buf)
398 {
399 	unsigned long result = 0;
400 
401 	while (n--) {
402 		result = (result << 32) | of_read_number(*buf, 1);
403 		(*buf)++;
404 	}
405 	return result;
406 }
407 
408 /*
409  * Read the next memblock list entry from the ibm,dynamic-memory property
410  * and return the information in the provided of_drconf_cell structure.
411  */
412 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
413 {
414 	const __be32 *cp;
415 
416 	drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
417 
418 	cp = *cellp;
419 	drmem->drc_index = of_read_number(cp, 1);
420 	drmem->reserved = of_read_number(&cp[1], 1);
421 	drmem->aa_index = of_read_number(&cp[2], 1);
422 	drmem->flags = of_read_number(&cp[3], 1);
423 
424 	*cellp = cp + 4;
425 }
426 
427 /*
428  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
429  *
430  * The layout of the ibm,dynamic-memory property is a number N of memblock
431  * list entries followed by N memblock list entries.  Each memblock list entry
432  * contains information as laid out in the of_drconf_cell struct above.
433  */
434 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
435 {
436 	const __be32 *prop;
437 	u32 len, entries;
438 
439 	prop = of_get_property(memory, "ibm,dynamic-memory", &len);
440 	if (!prop || len < sizeof(unsigned int))
441 		return 0;
442 
443 	entries = of_read_number(prop++, 1);
444 
445 	/* Now that we know the number of entries, revalidate the size
446 	 * of the property read in to ensure we have everything
447 	 */
448 	if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
449 		return 0;
450 
451 	*dm = prop;
452 	return entries;
453 }
454 
455 /*
456  * Retrieve and validate the ibm,lmb-size property for drconf memory
457  * from the device tree.
458  */
459 static u64 of_get_lmb_size(struct device_node *memory)
460 {
461 	const __be32 *prop;
462 	u32 len;
463 
464 	prop = of_get_property(memory, "ibm,lmb-size", &len);
465 	if (!prop || len < sizeof(unsigned int))
466 		return 0;
467 
468 	return read_n_cells(n_mem_size_cells, &prop);
469 }
470 
471 struct assoc_arrays {
472 	u32	n_arrays;
473 	u32	array_sz;
474 	const __be32 *arrays;
475 };
476 
477 /*
478  * Retrieve and validate the list of associativity arrays for drconf
479  * memory from the ibm,associativity-lookup-arrays property of the
480  * device tree..
481  *
482  * The layout of the ibm,associativity-lookup-arrays property is a number N
483  * indicating the number of associativity arrays, followed by a number M
484  * indicating the size of each associativity array, followed by a list
485  * of N associativity arrays.
486  */
487 static int of_get_assoc_arrays(struct device_node *memory,
488 			       struct assoc_arrays *aa)
489 {
490 	const __be32 *prop;
491 	u32 len;
492 
493 	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
494 	if (!prop || len < 2 * sizeof(unsigned int))
495 		return -1;
496 
497 	aa->n_arrays = of_read_number(prop++, 1);
498 	aa->array_sz = of_read_number(prop++, 1);
499 
500 	/* Now that we know the number of arrays and size of each array,
501 	 * revalidate the size of the property read in.
502 	 */
503 	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
504 		return -1;
505 
506 	aa->arrays = prop;
507 	return 0;
508 }
509 
510 /*
511  * This is like of_node_to_nid_single() for memory represented in the
512  * ibm,dynamic-reconfiguration-memory node.
513  */
514 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
515 				   struct assoc_arrays *aa)
516 {
517 	int default_nid = 0;
518 	int nid = default_nid;
519 	int index;
520 
521 	if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
522 	    !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
523 	    drmem->aa_index < aa->n_arrays) {
524 		index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
525 		nid = of_read_number(&aa->arrays[index], 1);
526 
527 		if (nid == 0xffff || nid >= MAX_NUMNODES)
528 			nid = default_nid;
529 	}
530 
531 	return nid;
532 }
533 
534 /*
535  * Figure out to which domain a cpu belongs and stick it there.
536  * Return the id of the domain used.
537  */
538 static int numa_setup_cpu(unsigned long lcpu)
539 {
540 	int nid;
541 	struct device_node *cpu;
542 
543 	/*
544 	 * If a valid cpu-to-node mapping is already available, use it
545 	 * directly instead of querying the firmware, since it represents
546 	 * the most recent mapping notified to us by the platform (eg: VPHN).
547 	 */
548 	if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
549 		map_cpu_to_node(lcpu, nid);
550 		return nid;
551 	}
552 
553 	cpu = of_get_cpu_node(lcpu, NULL);
554 
555 	if (!cpu) {
556 		WARN_ON(1);
557 		nid = 0;
558 		goto out;
559 	}
560 
561 	nid = of_node_to_nid_single(cpu);
562 
563 	if (nid < 0 || !node_online(nid))
564 		nid = first_online_node;
565 out:
566 	map_cpu_to_node(lcpu, nid);
567 
568 	of_node_put(cpu);
569 
570 	return nid;
571 }
572 
573 static void verify_cpu_node_mapping(int cpu, int node)
574 {
575 	int base, sibling, i;
576 
577 	/* Verify that all the threads in the core belong to the same node */
578 	base = cpu_first_thread_sibling(cpu);
579 
580 	for (i = 0; i < threads_per_core; i++) {
581 		sibling = base + i;
582 
583 		if (sibling == cpu || cpu_is_offline(sibling))
584 			continue;
585 
586 		if (cpu_to_node(sibling) != node) {
587 			WARN(1, "CPU thread siblings %d and %d don't belong"
588 				" to the same node!\n", cpu, sibling);
589 			break;
590 		}
591 	}
592 }
593 
594 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
595 			     void *hcpu)
596 {
597 	unsigned long lcpu = (unsigned long)hcpu;
598 	int ret = NOTIFY_DONE, nid;
599 
600 	switch (action) {
601 	case CPU_UP_PREPARE:
602 	case CPU_UP_PREPARE_FROZEN:
603 		nid = numa_setup_cpu(lcpu);
604 		verify_cpu_node_mapping((int)lcpu, nid);
605 		ret = NOTIFY_OK;
606 		break;
607 #ifdef CONFIG_HOTPLUG_CPU
608 	case CPU_DEAD:
609 	case CPU_DEAD_FROZEN:
610 	case CPU_UP_CANCELED:
611 	case CPU_UP_CANCELED_FROZEN:
612 		unmap_cpu_from_node(lcpu);
613 		break;
614 		ret = NOTIFY_OK;
615 #endif
616 	}
617 	return ret;
618 }
619 
620 /*
621  * Check and possibly modify a memory region to enforce the memory limit.
622  *
623  * Returns the size the region should have to enforce the memory limit.
624  * This will either be the original value of size, a truncated value,
625  * or zero. If the returned value of size is 0 the region should be
626  * discarded as it lies wholly above the memory limit.
627  */
628 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
629 						      unsigned long size)
630 {
631 	/*
632 	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
633 	 * we've already adjusted it for the limit and it takes care of
634 	 * having memory holes below the limit.  Also, in the case of
635 	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
636 	 */
637 
638 	if (start + size <= memblock_end_of_DRAM())
639 		return size;
640 
641 	if (start >= memblock_end_of_DRAM())
642 		return 0;
643 
644 	return memblock_end_of_DRAM() - start;
645 }
646 
647 /*
648  * Reads the counter for a given entry in
649  * linux,drconf-usable-memory property
650  */
651 static inline int __init read_usm_ranges(const __be32 **usm)
652 {
653 	/*
654 	 * For each lmb in ibm,dynamic-memory a corresponding
655 	 * entry in linux,drconf-usable-memory property contains
656 	 * a counter followed by that many (base, size) duple.
657 	 * read the counter from linux,drconf-usable-memory
658 	 */
659 	return read_n_cells(n_mem_size_cells, usm);
660 }
661 
662 /*
663  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
664  * node.  This assumes n_mem_{addr,size}_cells have been set.
665  */
666 static void __init parse_drconf_memory(struct device_node *memory)
667 {
668 	const __be32 *uninitialized_var(dm), *usm;
669 	unsigned int n, rc, ranges, is_kexec_kdump = 0;
670 	unsigned long lmb_size, base, size, sz;
671 	int nid;
672 	struct assoc_arrays aa = { .arrays = NULL };
673 
674 	n = of_get_drconf_memory(memory, &dm);
675 	if (!n)
676 		return;
677 
678 	lmb_size = of_get_lmb_size(memory);
679 	if (!lmb_size)
680 		return;
681 
682 	rc = of_get_assoc_arrays(memory, &aa);
683 	if (rc)
684 		return;
685 
686 	/* check if this is a kexec/kdump kernel */
687 	usm = of_get_usable_memory(memory);
688 	if (usm != NULL)
689 		is_kexec_kdump = 1;
690 
691 	for (; n != 0; --n) {
692 		struct of_drconf_cell drmem;
693 
694 		read_drconf_cell(&drmem, &dm);
695 
696 		/* skip this block if the reserved bit is set in flags (0x80)
697 		   or if the block is not assigned to this partition (0x8) */
698 		if ((drmem.flags & DRCONF_MEM_RESERVED)
699 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
700 			continue;
701 
702 		base = drmem.base_addr;
703 		size = lmb_size;
704 		ranges = 1;
705 
706 		if (is_kexec_kdump) {
707 			ranges = read_usm_ranges(&usm);
708 			if (!ranges) /* there are no (base, size) duple */
709 				continue;
710 		}
711 		do {
712 			if (is_kexec_kdump) {
713 				base = read_n_cells(n_mem_addr_cells, &usm);
714 				size = read_n_cells(n_mem_size_cells, &usm);
715 			}
716 			nid = of_drconf_to_nid_single(&drmem, &aa);
717 			fake_numa_create_new_node(
718 				((base + size) >> PAGE_SHIFT),
719 					   &nid);
720 			node_set_online(nid);
721 			sz = numa_enforce_memory_limit(base, size);
722 			if (sz)
723 				memblock_set_node(base, sz,
724 						  &memblock.memory, nid);
725 		} while (--ranges);
726 	}
727 }
728 
729 static int __init parse_numa_properties(void)
730 {
731 	struct device_node *memory;
732 	int default_nid = 0;
733 	unsigned long i;
734 
735 	if (numa_enabled == 0) {
736 		printk(KERN_WARNING "NUMA disabled by user\n");
737 		return -1;
738 	}
739 
740 	min_common_depth = find_min_common_depth();
741 
742 	if (min_common_depth < 0)
743 		return min_common_depth;
744 
745 	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
746 
747 	/*
748 	 * Even though we connect cpus to numa domains later in SMP
749 	 * init, we need to know the node ids now. This is because
750 	 * each node to be onlined must have NODE_DATA etc backing it.
751 	 */
752 	for_each_present_cpu(i) {
753 		struct device_node *cpu;
754 		int nid;
755 
756 		cpu = of_get_cpu_node(i, NULL);
757 		BUG_ON(!cpu);
758 		nid = of_node_to_nid_single(cpu);
759 		of_node_put(cpu);
760 
761 		/*
762 		 * Don't fall back to default_nid yet -- we will plug
763 		 * cpus into nodes once the memory scan has discovered
764 		 * the topology.
765 		 */
766 		if (nid < 0)
767 			continue;
768 		node_set_online(nid);
769 	}
770 
771 	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
772 
773 	for_each_node_by_type(memory, "memory") {
774 		unsigned long start;
775 		unsigned long size;
776 		int nid;
777 		int ranges;
778 		const __be32 *memcell_buf;
779 		unsigned int len;
780 
781 		memcell_buf = of_get_property(memory,
782 			"linux,usable-memory", &len);
783 		if (!memcell_buf || len <= 0)
784 			memcell_buf = of_get_property(memory, "reg", &len);
785 		if (!memcell_buf || len <= 0)
786 			continue;
787 
788 		/* ranges in cell */
789 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
790 new_range:
791 		/* these are order-sensitive, and modify the buffer pointer */
792 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
793 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
794 
795 		/*
796 		 * Assumption: either all memory nodes or none will
797 		 * have associativity properties.  If none, then
798 		 * everything goes to default_nid.
799 		 */
800 		nid = of_node_to_nid_single(memory);
801 		if (nid < 0)
802 			nid = default_nid;
803 
804 		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
805 		node_set_online(nid);
806 
807 		if (!(size = numa_enforce_memory_limit(start, size))) {
808 			if (--ranges)
809 				goto new_range;
810 			else
811 				continue;
812 		}
813 
814 		memblock_set_node(start, size, &memblock.memory, nid);
815 
816 		if (--ranges)
817 			goto new_range;
818 	}
819 
820 	/*
821 	 * Now do the same thing for each MEMBLOCK listed in the
822 	 * ibm,dynamic-memory property in the
823 	 * ibm,dynamic-reconfiguration-memory node.
824 	 */
825 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
826 	if (memory)
827 		parse_drconf_memory(memory);
828 
829 	return 0;
830 }
831 
832 static void __init setup_nonnuma(void)
833 {
834 	unsigned long top_of_ram = memblock_end_of_DRAM();
835 	unsigned long total_ram = memblock_phys_mem_size();
836 	unsigned long start_pfn, end_pfn;
837 	unsigned int nid = 0;
838 	struct memblock_region *reg;
839 
840 	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
841 	       top_of_ram, total_ram);
842 	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
843 	       (top_of_ram - total_ram) >> 20);
844 
845 	for_each_memblock(memory, reg) {
846 		start_pfn = memblock_region_memory_base_pfn(reg);
847 		end_pfn = memblock_region_memory_end_pfn(reg);
848 
849 		fake_numa_create_new_node(end_pfn, &nid);
850 		memblock_set_node(PFN_PHYS(start_pfn),
851 				  PFN_PHYS(end_pfn - start_pfn),
852 				  &memblock.memory, nid);
853 		node_set_online(nid);
854 	}
855 }
856 
857 void __init dump_numa_cpu_topology(void)
858 {
859 	unsigned int node;
860 	unsigned int cpu, count;
861 
862 	if (min_common_depth == -1 || !numa_enabled)
863 		return;
864 
865 	for_each_online_node(node) {
866 		printk(KERN_DEBUG "Node %d CPUs:", node);
867 
868 		count = 0;
869 		/*
870 		 * If we used a CPU iterator here we would miss printing
871 		 * the holes in the cpumap.
872 		 */
873 		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
874 			if (cpumask_test_cpu(cpu,
875 					node_to_cpumask_map[node])) {
876 				if (count == 0)
877 					printk(" %u", cpu);
878 				++count;
879 			} else {
880 				if (count > 1)
881 					printk("-%u", cpu - 1);
882 				count = 0;
883 			}
884 		}
885 
886 		if (count > 1)
887 			printk("-%u", nr_cpu_ids - 1);
888 		printk("\n");
889 	}
890 }
891 
892 static void __init dump_numa_memory_topology(void)
893 {
894 	unsigned int node;
895 	unsigned int count;
896 
897 	if (min_common_depth == -1 || !numa_enabled)
898 		return;
899 
900 	for_each_online_node(node) {
901 		unsigned long i;
902 
903 		printk(KERN_DEBUG "Node %d Memory:", node);
904 
905 		count = 0;
906 
907 		for (i = 0; i < memblock_end_of_DRAM();
908 		     i += (1 << SECTION_SIZE_BITS)) {
909 			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
910 				if (count == 0)
911 					printk(" 0x%lx", i);
912 				++count;
913 			} else {
914 				if (count > 0)
915 					printk("-0x%lx", i);
916 				count = 0;
917 			}
918 		}
919 
920 		if (count > 0)
921 			printk("-0x%lx", i);
922 		printk("\n");
923 	}
924 }
925 
926 /*
927  * Allocate some memory, satisfying the memblock or bootmem allocator where
928  * required. nid is the preferred node and end is the physical address of
929  * the highest address in the node.
930  *
931  * Returns the virtual address of the memory.
932  */
933 static void __init *careful_zallocation(int nid, unsigned long size,
934 				       unsigned long align,
935 				       unsigned long end_pfn)
936 {
937 	void *ret;
938 	int new_nid;
939 	unsigned long ret_paddr;
940 
941 	ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
942 
943 	/* retry over all memory */
944 	if (!ret_paddr)
945 		ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
946 
947 	if (!ret_paddr)
948 		panic("numa.c: cannot allocate %lu bytes for node %d",
949 		      size, nid);
950 
951 	ret = __va(ret_paddr);
952 
953 	/*
954 	 * We initialize the nodes in numeric order: 0, 1, 2...
955 	 * and hand over control from the MEMBLOCK allocator to the
956 	 * bootmem allocator.  If this function is called for
957 	 * node 5, then we know that all nodes <5 are using the
958 	 * bootmem allocator instead of the MEMBLOCK allocator.
959 	 *
960 	 * So, check the nid from which this allocation came
961 	 * and double check to see if we need to use bootmem
962 	 * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
963 	 * since it would be useless.
964 	 */
965 	new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
966 	if (new_nid < nid) {
967 		ret = __alloc_bootmem_node(NODE_DATA(new_nid),
968 				size, align, 0);
969 
970 		dbg("alloc_bootmem %p %lx\n", ret, size);
971 	}
972 
973 	memset(ret, 0, size);
974 	return ret;
975 }
976 
977 static struct notifier_block ppc64_numa_nb = {
978 	.notifier_call = cpu_numa_callback,
979 	.priority = 1 /* Must run before sched domains notifier. */
980 };
981 
982 static void __init mark_reserved_regions_for_nid(int nid)
983 {
984 	struct pglist_data *node = NODE_DATA(nid);
985 	struct memblock_region *reg;
986 
987 	for_each_memblock(reserved, reg) {
988 		unsigned long physbase = reg->base;
989 		unsigned long size = reg->size;
990 		unsigned long start_pfn = physbase >> PAGE_SHIFT;
991 		unsigned long end_pfn = PFN_UP(physbase + size);
992 		struct node_active_region node_ar;
993 		unsigned long node_end_pfn = pgdat_end_pfn(node);
994 
995 		/*
996 		 * Check to make sure that this memblock.reserved area is
997 		 * within the bounds of the node that we care about.
998 		 * Checking the nid of the start and end points is not
999 		 * sufficient because the reserved area could span the
1000 		 * entire node.
1001 		 */
1002 		if (end_pfn <= node->node_start_pfn ||
1003 		    start_pfn >= node_end_pfn)
1004 			continue;
1005 
1006 		get_node_active_region(start_pfn, &node_ar);
1007 		while (start_pfn < end_pfn &&
1008 			node_ar.start_pfn < node_ar.end_pfn) {
1009 			unsigned long reserve_size = size;
1010 			/*
1011 			 * if reserved region extends past active region
1012 			 * then trim size to active region
1013 			 */
1014 			if (end_pfn > node_ar.end_pfn)
1015 				reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1016 					- physbase;
1017 			/*
1018 			 * Only worry about *this* node, others may not
1019 			 * yet have valid NODE_DATA().
1020 			 */
1021 			if (node_ar.nid == nid) {
1022 				dbg("reserve_bootmem %lx %lx nid=%d\n",
1023 					physbase, reserve_size, node_ar.nid);
1024 				reserve_bootmem_node(NODE_DATA(node_ar.nid),
1025 						physbase, reserve_size,
1026 						BOOTMEM_DEFAULT);
1027 			}
1028 			/*
1029 			 * if reserved region is contained in the active region
1030 			 * then done.
1031 			 */
1032 			if (end_pfn <= node_ar.end_pfn)
1033 				break;
1034 
1035 			/*
1036 			 * reserved region extends past the active region
1037 			 *   get next active region that contains this
1038 			 *   reserved region
1039 			 */
1040 			start_pfn = node_ar.end_pfn;
1041 			physbase = start_pfn << PAGE_SHIFT;
1042 			size = size - reserve_size;
1043 			get_node_active_region(start_pfn, &node_ar);
1044 		}
1045 	}
1046 }
1047 
1048 
1049 void __init do_init_bootmem(void)
1050 {
1051 	int nid;
1052 
1053 	min_low_pfn = 0;
1054 	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1055 	max_pfn = max_low_pfn;
1056 
1057 	if (parse_numa_properties())
1058 		setup_nonnuma();
1059 	else
1060 		dump_numa_memory_topology();
1061 
1062 	for_each_online_node(nid) {
1063 		unsigned long start_pfn, end_pfn;
1064 		void *bootmem_vaddr;
1065 		unsigned long bootmap_pages;
1066 
1067 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1068 
1069 		/*
1070 		 * Allocate the node structure node local if possible
1071 		 *
1072 		 * Be careful moving this around, as it relies on all
1073 		 * previous nodes' bootmem to be initialized and have
1074 		 * all reserved areas marked.
1075 		 */
1076 		NODE_DATA(nid) = careful_zallocation(nid,
1077 					sizeof(struct pglist_data),
1078 					SMP_CACHE_BYTES, end_pfn);
1079 
1080   		dbg("node %d\n", nid);
1081 		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1082 
1083 		NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1084 		NODE_DATA(nid)->node_start_pfn = start_pfn;
1085 		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1086 
1087 		if (NODE_DATA(nid)->node_spanned_pages == 0)
1088   			continue;
1089 
1090   		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1091   		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1092 
1093 		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1094 		bootmem_vaddr = careful_zallocation(nid,
1095 					bootmap_pages << PAGE_SHIFT,
1096 					PAGE_SIZE, end_pfn);
1097 
1098 		dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1099 
1100 		init_bootmem_node(NODE_DATA(nid),
1101 				  __pa(bootmem_vaddr) >> PAGE_SHIFT,
1102 				  start_pfn, end_pfn);
1103 
1104 		free_bootmem_with_active_regions(nid, end_pfn);
1105 		/*
1106 		 * Be very careful about moving this around.  Future
1107 		 * calls to careful_zallocation() depend on this getting
1108 		 * done correctly.
1109 		 */
1110 		mark_reserved_regions_for_nid(nid);
1111 		sparse_memory_present_with_active_regions(nid);
1112 	}
1113 
1114 	init_bootmem_done = 1;
1115 
1116 	/*
1117 	 * Now bootmem is initialised we can create the node to cpumask
1118 	 * lookup tables and setup the cpu callback to populate them.
1119 	 */
1120 	setup_node_to_cpumask_map();
1121 
1122 	reset_numa_cpu_lookup_table();
1123 	register_cpu_notifier(&ppc64_numa_nb);
1124 	cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1125 			  (void *)(unsigned long)boot_cpuid);
1126 }
1127 
1128 void __init paging_init(void)
1129 {
1130 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1131 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1132 	max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1133 	free_area_init_nodes(max_zone_pfns);
1134 }
1135 
1136 static int __init early_numa(char *p)
1137 {
1138 	if (!p)
1139 		return 0;
1140 
1141 	if (strstr(p, "off"))
1142 		numa_enabled = 0;
1143 
1144 	if (strstr(p, "debug"))
1145 		numa_debug = 1;
1146 
1147 	p = strstr(p, "fake=");
1148 	if (p)
1149 		cmdline = p + strlen("fake=");
1150 
1151 	return 0;
1152 }
1153 early_param("numa", early_numa);
1154 
1155 #ifdef CONFIG_MEMORY_HOTPLUG
1156 /*
1157  * Find the node associated with a hot added memory section for
1158  * memory represented in the device tree by the property
1159  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1160  */
1161 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1162 				     unsigned long scn_addr)
1163 {
1164 	const __be32 *dm;
1165 	unsigned int drconf_cell_cnt, rc;
1166 	unsigned long lmb_size;
1167 	struct assoc_arrays aa;
1168 	int nid = -1;
1169 
1170 	drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1171 	if (!drconf_cell_cnt)
1172 		return -1;
1173 
1174 	lmb_size = of_get_lmb_size(memory);
1175 	if (!lmb_size)
1176 		return -1;
1177 
1178 	rc = of_get_assoc_arrays(memory, &aa);
1179 	if (rc)
1180 		return -1;
1181 
1182 	for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1183 		struct of_drconf_cell drmem;
1184 
1185 		read_drconf_cell(&drmem, &dm);
1186 
1187 		/* skip this block if it is reserved or not assigned to
1188 		 * this partition */
1189 		if ((drmem.flags & DRCONF_MEM_RESERVED)
1190 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1191 			continue;
1192 
1193 		if ((scn_addr < drmem.base_addr)
1194 		    || (scn_addr >= (drmem.base_addr + lmb_size)))
1195 			continue;
1196 
1197 		nid = of_drconf_to_nid_single(&drmem, &aa);
1198 		break;
1199 	}
1200 
1201 	return nid;
1202 }
1203 
1204 /*
1205  * Find the node associated with a hot added memory section for memory
1206  * represented in the device tree as a node (i.e. memory@XXXX) for
1207  * each memblock.
1208  */
1209 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1210 {
1211 	struct device_node *memory;
1212 	int nid = -1;
1213 
1214 	for_each_node_by_type(memory, "memory") {
1215 		unsigned long start, size;
1216 		int ranges;
1217 		const __be32 *memcell_buf;
1218 		unsigned int len;
1219 
1220 		memcell_buf = of_get_property(memory, "reg", &len);
1221 		if (!memcell_buf || len <= 0)
1222 			continue;
1223 
1224 		/* ranges in cell */
1225 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1226 
1227 		while (ranges--) {
1228 			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1229 			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1230 
1231 			if ((scn_addr < start) || (scn_addr >= (start + size)))
1232 				continue;
1233 
1234 			nid = of_node_to_nid_single(memory);
1235 			break;
1236 		}
1237 
1238 		if (nid >= 0)
1239 			break;
1240 	}
1241 
1242 	of_node_put(memory);
1243 
1244 	return nid;
1245 }
1246 
1247 /*
1248  * Find the node associated with a hot added memory section.  Section
1249  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1250  * sections are fully contained within a single MEMBLOCK.
1251  */
1252 int hot_add_scn_to_nid(unsigned long scn_addr)
1253 {
1254 	struct device_node *memory = NULL;
1255 	int nid, found = 0;
1256 
1257 	if (!numa_enabled || (min_common_depth < 0))
1258 		return first_online_node;
1259 
1260 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1261 	if (memory) {
1262 		nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1263 		of_node_put(memory);
1264 	} else {
1265 		nid = hot_add_node_scn_to_nid(scn_addr);
1266 	}
1267 
1268 	if (nid < 0 || !node_online(nid))
1269 		nid = first_online_node;
1270 
1271 	if (NODE_DATA(nid)->node_spanned_pages)
1272 		return nid;
1273 
1274 	for_each_online_node(nid) {
1275 		if (NODE_DATA(nid)->node_spanned_pages) {
1276 			found = 1;
1277 			break;
1278 		}
1279 	}
1280 
1281 	BUG_ON(!found);
1282 	return nid;
1283 }
1284 
1285 static u64 hot_add_drconf_memory_max(void)
1286 {
1287         struct device_node *memory = NULL;
1288         unsigned int drconf_cell_cnt = 0;
1289         u64 lmb_size = 0;
1290 	const __be32 *dm = NULL;
1291 
1292         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1293         if (memory) {
1294                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1295                 lmb_size = of_get_lmb_size(memory);
1296                 of_node_put(memory);
1297         }
1298         return lmb_size * drconf_cell_cnt;
1299 }
1300 
1301 /*
1302  * memory_hotplug_max - return max address of memory that may be added
1303  *
1304  * This is currently only used on systems that support drconfig memory
1305  * hotplug.
1306  */
1307 u64 memory_hotplug_max(void)
1308 {
1309         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1310 }
1311 #endif /* CONFIG_MEMORY_HOTPLUG */
1312 
1313 /* Virtual Processor Home Node (VPHN) support */
1314 #ifdef CONFIG_PPC_SPLPAR
1315 struct topology_update_data {
1316 	struct topology_update_data *next;
1317 	unsigned int cpu;
1318 	int old_nid;
1319 	int new_nid;
1320 };
1321 
1322 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1323 static cpumask_t cpu_associativity_changes_mask;
1324 static int vphn_enabled;
1325 static int prrn_enabled;
1326 static void reset_topology_timer(void);
1327 
1328 /*
1329  * Store the current values of the associativity change counters in the
1330  * hypervisor.
1331  */
1332 static void setup_cpu_associativity_change_counters(void)
1333 {
1334 	int cpu;
1335 
1336 	/* The VPHN feature supports a maximum of 8 reference points */
1337 	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1338 
1339 	for_each_possible_cpu(cpu) {
1340 		int i;
1341 		u8 *counts = vphn_cpu_change_counts[cpu];
1342 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1343 
1344 		for (i = 0; i < distance_ref_points_depth; i++)
1345 			counts[i] = hypervisor_counts[i];
1346 	}
1347 }
1348 
1349 /*
1350  * The hypervisor maintains a set of 8 associativity change counters in
1351  * the VPA of each cpu that correspond to the associativity levels in the
1352  * ibm,associativity-reference-points property. When an associativity
1353  * level changes, the corresponding counter is incremented.
1354  *
1355  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1356  * node associativity levels have changed.
1357  *
1358  * Returns the number of cpus with unhandled associativity changes.
1359  */
1360 static int update_cpu_associativity_changes_mask(void)
1361 {
1362 	int cpu;
1363 	cpumask_t *changes = &cpu_associativity_changes_mask;
1364 
1365 	for_each_possible_cpu(cpu) {
1366 		int i, changed = 0;
1367 		u8 *counts = vphn_cpu_change_counts[cpu];
1368 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1369 
1370 		for (i = 0; i < distance_ref_points_depth; i++) {
1371 			if (hypervisor_counts[i] != counts[i]) {
1372 				counts[i] = hypervisor_counts[i];
1373 				changed = 1;
1374 			}
1375 		}
1376 		if (changed) {
1377 			cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1378 			cpu = cpu_last_thread_sibling(cpu);
1379 		}
1380 	}
1381 
1382 	return cpumask_weight(changes);
1383 }
1384 
1385 /*
1386  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1387  * the complete property we have to add the length in the first cell.
1388  */
1389 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1390 
1391 /*
1392  * Convert the associativity domain numbers returned from the hypervisor
1393  * to the sequence they would appear in the ibm,associativity property.
1394  */
1395 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1396 {
1397 	int i, nr_assoc_doms = 0;
1398 	const __be16 *field = (const __be16 *) packed;
1399 
1400 #define VPHN_FIELD_UNUSED	(0xffff)
1401 #define VPHN_FIELD_MSB		(0x8000)
1402 #define VPHN_FIELD_MASK		(~VPHN_FIELD_MSB)
1403 
1404 	for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1405 		if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1406 			/* All significant fields processed, and remaining
1407 			 * fields contain the reserved value of all 1's.
1408 			 * Just store them.
1409 			 */
1410 			unpacked[i] = *((__be32 *)field);
1411 			field += 2;
1412 		} else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1413 			/* Data is in the lower 15 bits of this field */
1414 			unpacked[i] = cpu_to_be32(
1415 				be16_to_cpup(field) & VPHN_FIELD_MASK);
1416 			field++;
1417 			nr_assoc_doms++;
1418 		} else {
1419 			/* Data is in the lower 15 bits of this field
1420 			 * concatenated with the next 16 bit field
1421 			 */
1422 			unpacked[i] = *((__be32 *)field);
1423 			field += 2;
1424 			nr_assoc_doms++;
1425 		}
1426 	}
1427 
1428 	/* The first cell contains the length of the property */
1429 	unpacked[0] = cpu_to_be32(nr_assoc_doms);
1430 
1431 	return nr_assoc_doms;
1432 }
1433 
1434 /*
1435  * Retrieve the new associativity information for a virtual processor's
1436  * home node.
1437  */
1438 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1439 {
1440 	long rc;
1441 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1442 	u64 flags = 1;
1443 	int hwcpu = get_hard_smp_processor_id(cpu);
1444 
1445 	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1446 	vphn_unpack_associativity(retbuf, associativity);
1447 
1448 	return rc;
1449 }
1450 
1451 static long vphn_get_associativity(unsigned long cpu,
1452 					__be32 *associativity)
1453 {
1454 	long rc;
1455 
1456 	rc = hcall_vphn(cpu, associativity);
1457 
1458 	switch (rc) {
1459 	case H_FUNCTION:
1460 		printk(KERN_INFO
1461 			"VPHN is not supported. Disabling polling...\n");
1462 		stop_topology_update();
1463 		break;
1464 	case H_HARDWARE:
1465 		printk(KERN_ERR
1466 			"hcall_vphn() experienced a hardware fault "
1467 			"preventing VPHN. Disabling polling...\n");
1468 		stop_topology_update();
1469 	}
1470 
1471 	return rc;
1472 }
1473 
1474 /*
1475  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1476  * characteristics change. This function doesn't perform any locking and is
1477  * only safe to call from stop_machine().
1478  */
1479 static int update_cpu_topology(void *data)
1480 {
1481 	struct topology_update_data *update;
1482 	unsigned long cpu;
1483 
1484 	if (!data)
1485 		return -EINVAL;
1486 
1487 	cpu = smp_processor_id();
1488 
1489 	for (update = data; update; update = update->next) {
1490 		if (cpu != update->cpu)
1491 			continue;
1492 
1493 		unmap_cpu_from_node(update->cpu);
1494 		map_cpu_to_node(update->cpu, update->new_nid);
1495 		vdso_getcpu_init();
1496 	}
1497 
1498 	return 0;
1499 }
1500 
1501 static int update_lookup_table(void *data)
1502 {
1503 	struct topology_update_data *update;
1504 
1505 	if (!data)
1506 		return -EINVAL;
1507 
1508 	/*
1509 	 * Upon topology update, the numa-cpu lookup table needs to be updated
1510 	 * for all threads in the core, including offline CPUs, to ensure that
1511 	 * future hotplug operations respect the cpu-to-node associativity
1512 	 * properly.
1513 	 */
1514 	for (update = data; update; update = update->next) {
1515 		int nid, base, j;
1516 
1517 		nid = update->new_nid;
1518 		base = cpu_first_thread_sibling(update->cpu);
1519 
1520 		for (j = 0; j < threads_per_core; j++) {
1521 			update_numa_cpu_lookup_table(base + j, nid);
1522 		}
1523 	}
1524 
1525 	return 0;
1526 }
1527 
1528 /*
1529  * Update the node maps and sysfs entries for each cpu whose home node
1530  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1531  */
1532 int arch_update_cpu_topology(void)
1533 {
1534 	unsigned int cpu, sibling, changed = 0;
1535 	struct topology_update_data *updates, *ud;
1536 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1537 	cpumask_t updated_cpus;
1538 	struct device *dev;
1539 	int weight, new_nid, i = 0;
1540 
1541 	weight = cpumask_weight(&cpu_associativity_changes_mask);
1542 	if (!weight)
1543 		return 0;
1544 
1545 	updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1546 	if (!updates)
1547 		return 0;
1548 
1549 	cpumask_clear(&updated_cpus);
1550 
1551 	for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1552 		/*
1553 		 * If siblings aren't flagged for changes, updates list
1554 		 * will be too short. Skip on this update and set for next
1555 		 * update.
1556 		 */
1557 		if (!cpumask_subset(cpu_sibling_mask(cpu),
1558 					&cpu_associativity_changes_mask)) {
1559 			pr_info("Sibling bits not set for associativity "
1560 					"change, cpu%d\n", cpu);
1561 			cpumask_or(&cpu_associativity_changes_mask,
1562 					&cpu_associativity_changes_mask,
1563 					cpu_sibling_mask(cpu));
1564 			cpu = cpu_last_thread_sibling(cpu);
1565 			continue;
1566 		}
1567 
1568 		/* Use associativity from first thread for all siblings */
1569 		vphn_get_associativity(cpu, associativity);
1570 		new_nid = associativity_to_nid(associativity);
1571 		if (new_nid < 0 || !node_online(new_nid))
1572 			new_nid = first_online_node;
1573 
1574 		if (new_nid == numa_cpu_lookup_table[cpu]) {
1575 			cpumask_andnot(&cpu_associativity_changes_mask,
1576 					&cpu_associativity_changes_mask,
1577 					cpu_sibling_mask(cpu));
1578 			cpu = cpu_last_thread_sibling(cpu);
1579 			continue;
1580 		}
1581 
1582 		for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1583 			ud = &updates[i++];
1584 			ud->cpu = sibling;
1585 			ud->new_nid = new_nid;
1586 			ud->old_nid = numa_cpu_lookup_table[sibling];
1587 			cpumask_set_cpu(sibling, &updated_cpus);
1588 			if (i < weight)
1589 				ud->next = &updates[i];
1590 		}
1591 		cpu = cpu_last_thread_sibling(cpu);
1592 	}
1593 
1594 	/*
1595 	 * In cases where we have nothing to update (because the updates list
1596 	 * is too short or because the new topology is same as the old one),
1597 	 * skip invoking update_cpu_topology() via stop-machine(). This is
1598 	 * necessary (and not just a fast-path optimization) since stop-machine
1599 	 * can end up electing a random CPU to run update_cpu_topology(), and
1600 	 * thus trick us into setting up incorrect cpu-node mappings (since
1601 	 * 'updates' is kzalloc()'ed).
1602 	 *
1603 	 * And for the similar reason, we will skip all the following updating.
1604 	 */
1605 	if (!cpumask_weight(&updated_cpus))
1606 		goto out;
1607 
1608 	stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1609 
1610 	/*
1611 	 * Update the numa-cpu lookup table with the new mappings, even for
1612 	 * offline CPUs. It is best to perform this update from the stop-
1613 	 * machine context.
1614 	 */
1615 	stop_machine(update_lookup_table, &updates[0],
1616 					cpumask_of(raw_smp_processor_id()));
1617 
1618 	for (ud = &updates[0]; ud; ud = ud->next) {
1619 		unregister_cpu_under_node(ud->cpu, ud->old_nid);
1620 		register_cpu_under_node(ud->cpu, ud->new_nid);
1621 
1622 		dev = get_cpu_device(ud->cpu);
1623 		if (dev)
1624 			kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1625 		cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1626 		changed = 1;
1627 	}
1628 
1629 out:
1630 	kfree(updates);
1631 	return changed;
1632 }
1633 
1634 static void topology_work_fn(struct work_struct *work)
1635 {
1636 	rebuild_sched_domains();
1637 }
1638 static DECLARE_WORK(topology_work, topology_work_fn);
1639 
1640 static void topology_schedule_update(void)
1641 {
1642 	schedule_work(&topology_work);
1643 }
1644 
1645 static void topology_timer_fn(unsigned long ignored)
1646 {
1647 	if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1648 		topology_schedule_update();
1649 	else if (vphn_enabled) {
1650 		if (update_cpu_associativity_changes_mask() > 0)
1651 			topology_schedule_update();
1652 		reset_topology_timer();
1653 	}
1654 }
1655 static struct timer_list topology_timer =
1656 	TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1657 
1658 static void reset_topology_timer(void)
1659 {
1660 	topology_timer.data = 0;
1661 	topology_timer.expires = jiffies + 60 * HZ;
1662 	mod_timer(&topology_timer, topology_timer.expires);
1663 }
1664 
1665 #ifdef CONFIG_SMP
1666 
1667 static void stage_topology_update(int core_id)
1668 {
1669 	cpumask_or(&cpu_associativity_changes_mask,
1670 		&cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1671 	reset_topology_timer();
1672 }
1673 
1674 static int dt_update_callback(struct notifier_block *nb,
1675 				unsigned long action, void *data)
1676 {
1677 	struct of_prop_reconfig *update;
1678 	int rc = NOTIFY_DONE;
1679 
1680 	switch (action) {
1681 	case OF_RECONFIG_UPDATE_PROPERTY:
1682 		update = (struct of_prop_reconfig *)data;
1683 		if (!of_prop_cmp(update->dn->type, "cpu") &&
1684 		    !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1685 			u32 core_id;
1686 			of_property_read_u32(update->dn, "reg", &core_id);
1687 			stage_topology_update(core_id);
1688 			rc = NOTIFY_OK;
1689 		}
1690 		break;
1691 	}
1692 
1693 	return rc;
1694 }
1695 
1696 static struct notifier_block dt_update_nb = {
1697 	.notifier_call = dt_update_callback,
1698 };
1699 
1700 #endif
1701 
1702 /*
1703  * Start polling for associativity changes.
1704  */
1705 int start_topology_update(void)
1706 {
1707 	int rc = 0;
1708 
1709 	if (firmware_has_feature(FW_FEATURE_PRRN)) {
1710 		if (!prrn_enabled) {
1711 			prrn_enabled = 1;
1712 			vphn_enabled = 0;
1713 #ifdef CONFIG_SMP
1714 			rc = of_reconfig_notifier_register(&dt_update_nb);
1715 #endif
1716 		}
1717 	} else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1718 		   lppaca_shared_proc(get_lppaca())) {
1719 		if (!vphn_enabled) {
1720 			prrn_enabled = 0;
1721 			vphn_enabled = 1;
1722 			setup_cpu_associativity_change_counters();
1723 			init_timer_deferrable(&topology_timer);
1724 			reset_topology_timer();
1725 		}
1726 	}
1727 
1728 	return rc;
1729 }
1730 
1731 /*
1732  * Disable polling for VPHN associativity changes.
1733  */
1734 int stop_topology_update(void)
1735 {
1736 	int rc = 0;
1737 
1738 	if (prrn_enabled) {
1739 		prrn_enabled = 0;
1740 #ifdef CONFIG_SMP
1741 		rc = of_reconfig_notifier_unregister(&dt_update_nb);
1742 #endif
1743 	} else if (vphn_enabled) {
1744 		vphn_enabled = 0;
1745 		rc = del_timer_sync(&topology_timer);
1746 	}
1747 
1748 	return rc;
1749 }
1750 
1751 int prrn_is_enabled(void)
1752 {
1753 	return prrn_enabled;
1754 }
1755 
1756 static int topology_read(struct seq_file *file, void *v)
1757 {
1758 	if (vphn_enabled || prrn_enabled)
1759 		seq_puts(file, "on\n");
1760 	else
1761 		seq_puts(file, "off\n");
1762 
1763 	return 0;
1764 }
1765 
1766 static int topology_open(struct inode *inode, struct file *file)
1767 {
1768 	return single_open(file, topology_read, NULL);
1769 }
1770 
1771 static ssize_t topology_write(struct file *file, const char __user *buf,
1772 			      size_t count, loff_t *off)
1773 {
1774 	char kbuf[4]; /* "on" or "off" plus null. */
1775 	int read_len;
1776 
1777 	read_len = count < 3 ? count : 3;
1778 	if (copy_from_user(kbuf, buf, read_len))
1779 		return -EINVAL;
1780 
1781 	kbuf[read_len] = '\0';
1782 
1783 	if (!strncmp(kbuf, "on", 2))
1784 		start_topology_update();
1785 	else if (!strncmp(kbuf, "off", 3))
1786 		stop_topology_update();
1787 	else
1788 		return -EINVAL;
1789 
1790 	return count;
1791 }
1792 
1793 static const struct file_operations topology_ops = {
1794 	.read = seq_read,
1795 	.write = topology_write,
1796 	.open = topology_open,
1797 	.release = single_release
1798 };
1799 
1800 static int topology_update_init(void)
1801 {
1802 	start_topology_update();
1803 	proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops);
1804 
1805 	return 0;
1806 }
1807 device_initcall(topology_update_init);
1808 #endif /* CONFIG_PPC_SPLPAR */
1809