xref: /openbmc/linux/arch/powerpc/mm/numa.c (revision 206a81c1)
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 EXPORT_SYMBOL(__node_distance);
236 
237 static void initialize_distance_lookup_table(int nid,
238 		const __be32 *associativity)
239 {
240 	int i;
241 
242 	if (!form1_affinity)
243 		return;
244 
245 	for (i = 0; i < distance_ref_points_depth; i++) {
246 		const __be32 *entry;
247 
248 		entry = &associativity[be32_to_cpu(distance_ref_points[i])];
249 		distance_lookup_table[nid][i] = of_read_number(entry, 1);
250 	}
251 }
252 
253 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
254  * info is found.
255  */
256 static int associativity_to_nid(const __be32 *associativity)
257 {
258 	int nid = -1;
259 
260 	if (min_common_depth == -1)
261 		goto out;
262 
263 	if (of_read_number(associativity, 1) >= min_common_depth)
264 		nid = of_read_number(&associativity[min_common_depth], 1);
265 
266 	/* POWER4 LPAR uses 0xffff as invalid node */
267 	if (nid == 0xffff || nid >= MAX_NUMNODES)
268 		nid = -1;
269 
270 	if (nid > 0 &&
271 	    of_read_number(associativity, 1) >= distance_ref_points_depth)
272 		initialize_distance_lookup_table(nid, associativity);
273 
274 out:
275 	return nid;
276 }
277 
278 /* Returns the nid associated with the given device tree node,
279  * or -1 if not found.
280  */
281 static int of_node_to_nid_single(struct device_node *device)
282 {
283 	int nid = -1;
284 	const __be32 *tmp;
285 
286 	tmp = of_get_associativity(device);
287 	if (tmp)
288 		nid = associativity_to_nid(tmp);
289 	return nid;
290 }
291 
292 /* Walk the device tree upwards, looking for an associativity id */
293 int of_node_to_nid(struct device_node *device)
294 {
295 	struct device_node *tmp;
296 	int nid = -1;
297 
298 	of_node_get(device);
299 	while (device) {
300 		nid = of_node_to_nid_single(device);
301 		if (nid != -1)
302 			break;
303 
304 	        tmp = device;
305 		device = of_get_parent(tmp);
306 		of_node_put(tmp);
307 	}
308 	of_node_put(device);
309 
310 	return nid;
311 }
312 EXPORT_SYMBOL_GPL(of_node_to_nid);
313 
314 static int __init find_min_common_depth(void)
315 {
316 	int depth;
317 	struct device_node *root;
318 
319 	if (firmware_has_feature(FW_FEATURE_OPAL))
320 		root = of_find_node_by_path("/ibm,opal");
321 	else
322 		root = of_find_node_by_path("/rtas");
323 	if (!root)
324 		root = of_find_node_by_path("/");
325 
326 	/*
327 	 * This property is a set of 32-bit integers, each representing
328 	 * an index into the ibm,associativity nodes.
329 	 *
330 	 * With form 0 affinity the first integer is for an SMP configuration
331 	 * (should be all 0's) and the second is for a normal NUMA
332 	 * configuration. We have only one level of NUMA.
333 	 *
334 	 * With form 1 affinity the first integer is the most significant
335 	 * NUMA boundary and the following are progressively less significant
336 	 * boundaries. There can be more than one level of NUMA.
337 	 */
338 	distance_ref_points = of_get_property(root,
339 					"ibm,associativity-reference-points",
340 					&distance_ref_points_depth);
341 
342 	if (!distance_ref_points) {
343 		dbg("NUMA: ibm,associativity-reference-points not found.\n");
344 		goto err;
345 	}
346 
347 	distance_ref_points_depth /= sizeof(int);
348 
349 	if (firmware_has_feature(FW_FEATURE_OPAL) ||
350 	    firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
351 		dbg("Using form 1 affinity\n");
352 		form1_affinity = 1;
353 	}
354 
355 	if (form1_affinity) {
356 		depth = of_read_number(distance_ref_points, 1);
357 	} else {
358 		if (distance_ref_points_depth < 2) {
359 			printk(KERN_WARNING "NUMA: "
360 				"short ibm,associativity-reference-points\n");
361 			goto err;
362 		}
363 
364 		depth = of_read_number(&distance_ref_points[1], 1);
365 	}
366 
367 	/*
368 	 * Warn and cap if the hardware supports more than
369 	 * MAX_DISTANCE_REF_POINTS domains.
370 	 */
371 	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
372 		printk(KERN_WARNING "NUMA: distance array capped at "
373 			"%d entries\n", MAX_DISTANCE_REF_POINTS);
374 		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
375 	}
376 
377 	of_node_put(root);
378 	return depth;
379 
380 err:
381 	of_node_put(root);
382 	return -1;
383 }
384 
385 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
386 {
387 	struct device_node *memory = NULL;
388 
389 	memory = of_find_node_by_type(memory, "memory");
390 	if (!memory)
391 		panic("numa.c: No memory nodes found!");
392 
393 	*n_addr_cells = of_n_addr_cells(memory);
394 	*n_size_cells = of_n_size_cells(memory);
395 	of_node_put(memory);
396 }
397 
398 static unsigned long read_n_cells(int n, const __be32 **buf)
399 {
400 	unsigned long result = 0;
401 
402 	while (n--) {
403 		result = (result << 32) | of_read_number(*buf, 1);
404 		(*buf)++;
405 	}
406 	return result;
407 }
408 
409 /*
410  * Read the next memblock list entry from the ibm,dynamic-memory property
411  * and return the information in the provided of_drconf_cell structure.
412  */
413 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
414 {
415 	const __be32 *cp;
416 
417 	drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
418 
419 	cp = *cellp;
420 	drmem->drc_index = of_read_number(cp, 1);
421 	drmem->reserved = of_read_number(&cp[1], 1);
422 	drmem->aa_index = of_read_number(&cp[2], 1);
423 	drmem->flags = of_read_number(&cp[3], 1);
424 
425 	*cellp = cp + 4;
426 }
427 
428 /*
429  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
430  *
431  * The layout of the ibm,dynamic-memory property is a number N of memblock
432  * list entries followed by N memblock list entries.  Each memblock list entry
433  * contains information as laid out in the of_drconf_cell struct above.
434  */
435 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
436 {
437 	const __be32 *prop;
438 	u32 len, entries;
439 
440 	prop = of_get_property(memory, "ibm,dynamic-memory", &len);
441 	if (!prop || len < sizeof(unsigned int))
442 		return 0;
443 
444 	entries = of_read_number(prop++, 1);
445 
446 	/* Now that we know the number of entries, revalidate the size
447 	 * of the property read in to ensure we have everything
448 	 */
449 	if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
450 		return 0;
451 
452 	*dm = prop;
453 	return entries;
454 }
455 
456 /*
457  * Retrieve and validate the ibm,lmb-size property for drconf memory
458  * from the device tree.
459  */
460 static u64 of_get_lmb_size(struct device_node *memory)
461 {
462 	const __be32 *prop;
463 	u32 len;
464 
465 	prop = of_get_property(memory, "ibm,lmb-size", &len);
466 	if (!prop || len < sizeof(unsigned int))
467 		return 0;
468 
469 	return read_n_cells(n_mem_size_cells, &prop);
470 }
471 
472 struct assoc_arrays {
473 	u32	n_arrays;
474 	u32	array_sz;
475 	const __be32 *arrays;
476 };
477 
478 /*
479  * Retrieve and validate the list of associativity arrays for drconf
480  * memory from the ibm,associativity-lookup-arrays property of the
481  * device tree..
482  *
483  * The layout of the ibm,associativity-lookup-arrays property is a number N
484  * indicating the number of associativity arrays, followed by a number M
485  * indicating the size of each associativity array, followed by a list
486  * of N associativity arrays.
487  */
488 static int of_get_assoc_arrays(struct device_node *memory,
489 			       struct assoc_arrays *aa)
490 {
491 	const __be32 *prop;
492 	u32 len;
493 
494 	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
495 	if (!prop || len < 2 * sizeof(unsigned int))
496 		return -1;
497 
498 	aa->n_arrays = of_read_number(prop++, 1);
499 	aa->array_sz = of_read_number(prop++, 1);
500 
501 	/* Now that we know the number of arrays and size of each array,
502 	 * revalidate the size of the property read in.
503 	 */
504 	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
505 		return -1;
506 
507 	aa->arrays = prop;
508 	return 0;
509 }
510 
511 /*
512  * This is like of_node_to_nid_single() for memory represented in the
513  * ibm,dynamic-reconfiguration-memory node.
514  */
515 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
516 				   struct assoc_arrays *aa)
517 {
518 	int default_nid = 0;
519 	int nid = default_nid;
520 	int index;
521 
522 	if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
523 	    !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
524 	    drmem->aa_index < aa->n_arrays) {
525 		index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
526 		nid = of_read_number(&aa->arrays[index], 1);
527 
528 		if (nid == 0xffff || nid >= MAX_NUMNODES)
529 			nid = default_nid;
530 	}
531 
532 	return nid;
533 }
534 
535 /*
536  * Figure out to which domain a cpu belongs and stick it there.
537  * Return the id of the domain used.
538  */
539 static int numa_setup_cpu(unsigned long lcpu)
540 {
541 	int nid;
542 	struct device_node *cpu;
543 
544 	/*
545 	 * If a valid cpu-to-node mapping is already available, use it
546 	 * directly instead of querying the firmware, since it represents
547 	 * the most recent mapping notified to us by the platform (eg: VPHN).
548 	 */
549 	if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
550 		map_cpu_to_node(lcpu, nid);
551 		return nid;
552 	}
553 
554 	cpu = of_get_cpu_node(lcpu, NULL);
555 
556 	if (!cpu) {
557 		WARN_ON(1);
558 		nid = 0;
559 		goto out;
560 	}
561 
562 	nid = of_node_to_nid_single(cpu);
563 
564 	if (nid < 0 || !node_online(nid))
565 		nid = first_online_node;
566 out:
567 	map_cpu_to_node(lcpu, nid);
568 
569 	of_node_put(cpu);
570 
571 	return nid;
572 }
573 
574 static void verify_cpu_node_mapping(int cpu, int node)
575 {
576 	int base, sibling, i;
577 
578 	/* Verify that all the threads in the core belong to the same node */
579 	base = cpu_first_thread_sibling(cpu);
580 
581 	for (i = 0; i < threads_per_core; i++) {
582 		sibling = base + i;
583 
584 		if (sibling == cpu || cpu_is_offline(sibling))
585 			continue;
586 
587 		if (cpu_to_node(sibling) != node) {
588 			WARN(1, "CPU thread siblings %d and %d don't belong"
589 				" to the same node!\n", cpu, sibling);
590 			break;
591 		}
592 	}
593 }
594 
595 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
596 			     void *hcpu)
597 {
598 	unsigned long lcpu = (unsigned long)hcpu;
599 	int ret = NOTIFY_DONE, nid;
600 
601 	switch (action) {
602 	case CPU_UP_PREPARE:
603 	case CPU_UP_PREPARE_FROZEN:
604 		nid = numa_setup_cpu(lcpu);
605 		verify_cpu_node_mapping((int)lcpu, nid);
606 		ret = NOTIFY_OK;
607 		break;
608 #ifdef CONFIG_HOTPLUG_CPU
609 	case CPU_DEAD:
610 	case CPU_DEAD_FROZEN:
611 	case CPU_UP_CANCELED:
612 	case CPU_UP_CANCELED_FROZEN:
613 		unmap_cpu_from_node(lcpu);
614 		break;
615 		ret = NOTIFY_OK;
616 #endif
617 	}
618 	return ret;
619 }
620 
621 /*
622  * Check and possibly modify a memory region to enforce the memory limit.
623  *
624  * Returns the size the region should have to enforce the memory limit.
625  * This will either be the original value of size, a truncated value,
626  * or zero. If the returned value of size is 0 the region should be
627  * discarded as it lies wholly above the memory limit.
628  */
629 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
630 						      unsigned long size)
631 {
632 	/*
633 	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
634 	 * we've already adjusted it for the limit and it takes care of
635 	 * having memory holes below the limit.  Also, in the case of
636 	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
637 	 */
638 
639 	if (start + size <= memblock_end_of_DRAM())
640 		return size;
641 
642 	if (start >= memblock_end_of_DRAM())
643 		return 0;
644 
645 	return memblock_end_of_DRAM() - start;
646 }
647 
648 /*
649  * Reads the counter for a given entry in
650  * linux,drconf-usable-memory property
651  */
652 static inline int __init read_usm_ranges(const __be32 **usm)
653 {
654 	/*
655 	 * For each lmb in ibm,dynamic-memory a corresponding
656 	 * entry in linux,drconf-usable-memory property contains
657 	 * a counter followed by that many (base, size) duple.
658 	 * read the counter from linux,drconf-usable-memory
659 	 */
660 	return read_n_cells(n_mem_size_cells, usm);
661 }
662 
663 /*
664  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
665  * node.  This assumes n_mem_{addr,size}_cells have been set.
666  */
667 static void __init parse_drconf_memory(struct device_node *memory)
668 {
669 	const __be32 *uninitialized_var(dm), *usm;
670 	unsigned int n, rc, ranges, is_kexec_kdump = 0;
671 	unsigned long lmb_size, base, size, sz;
672 	int nid;
673 	struct assoc_arrays aa = { .arrays = NULL };
674 
675 	n = of_get_drconf_memory(memory, &dm);
676 	if (!n)
677 		return;
678 
679 	lmb_size = of_get_lmb_size(memory);
680 	if (!lmb_size)
681 		return;
682 
683 	rc = of_get_assoc_arrays(memory, &aa);
684 	if (rc)
685 		return;
686 
687 	/* check if this is a kexec/kdump kernel */
688 	usm = of_get_usable_memory(memory);
689 	if (usm != NULL)
690 		is_kexec_kdump = 1;
691 
692 	for (; n != 0; --n) {
693 		struct of_drconf_cell drmem;
694 
695 		read_drconf_cell(&drmem, &dm);
696 
697 		/* skip this block if the reserved bit is set in flags (0x80)
698 		   or if the block is not assigned to this partition (0x8) */
699 		if ((drmem.flags & DRCONF_MEM_RESERVED)
700 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
701 			continue;
702 
703 		base = drmem.base_addr;
704 		size = lmb_size;
705 		ranges = 1;
706 
707 		if (is_kexec_kdump) {
708 			ranges = read_usm_ranges(&usm);
709 			if (!ranges) /* there are no (base, size) duple */
710 				continue;
711 		}
712 		do {
713 			if (is_kexec_kdump) {
714 				base = read_n_cells(n_mem_addr_cells, &usm);
715 				size = read_n_cells(n_mem_size_cells, &usm);
716 			}
717 			nid = of_drconf_to_nid_single(&drmem, &aa);
718 			fake_numa_create_new_node(
719 				((base + size) >> PAGE_SHIFT),
720 					   &nid);
721 			node_set_online(nid);
722 			sz = numa_enforce_memory_limit(base, size);
723 			if (sz)
724 				memblock_set_node(base, sz,
725 						  &memblock.memory, nid);
726 		} while (--ranges);
727 	}
728 }
729 
730 static int __init parse_numa_properties(void)
731 {
732 	struct device_node *memory;
733 	int default_nid = 0;
734 	unsigned long i;
735 
736 	if (numa_enabled == 0) {
737 		printk(KERN_WARNING "NUMA disabled by user\n");
738 		return -1;
739 	}
740 
741 	min_common_depth = find_min_common_depth();
742 
743 	if (min_common_depth < 0)
744 		return min_common_depth;
745 
746 	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
747 
748 	/*
749 	 * Even though we connect cpus to numa domains later in SMP
750 	 * init, we need to know the node ids now. This is because
751 	 * each node to be onlined must have NODE_DATA etc backing it.
752 	 */
753 	for_each_present_cpu(i) {
754 		struct device_node *cpu;
755 		int nid;
756 
757 		cpu = of_get_cpu_node(i, NULL);
758 		BUG_ON(!cpu);
759 		nid = of_node_to_nid_single(cpu);
760 		of_node_put(cpu);
761 
762 		/*
763 		 * Don't fall back to default_nid yet -- we will plug
764 		 * cpus into nodes once the memory scan has discovered
765 		 * the topology.
766 		 */
767 		if (nid < 0)
768 			continue;
769 		node_set_online(nid);
770 	}
771 
772 	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
773 
774 	for_each_node_by_type(memory, "memory") {
775 		unsigned long start;
776 		unsigned long size;
777 		int nid;
778 		int ranges;
779 		const __be32 *memcell_buf;
780 		unsigned int len;
781 
782 		memcell_buf = of_get_property(memory,
783 			"linux,usable-memory", &len);
784 		if (!memcell_buf || len <= 0)
785 			memcell_buf = of_get_property(memory, "reg", &len);
786 		if (!memcell_buf || len <= 0)
787 			continue;
788 
789 		/* ranges in cell */
790 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
791 new_range:
792 		/* these are order-sensitive, and modify the buffer pointer */
793 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
794 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
795 
796 		/*
797 		 * Assumption: either all memory nodes or none will
798 		 * have associativity properties.  If none, then
799 		 * everything goes to default_nid.
800 		 */
801 		nid = of_node_to_nid_single(memory);
802 		if (nid < 0)
803 			nid = default_nid;
804 
805 		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
806 		node_set_online(nid);
807 
808 		if (!(size = numa_enforce_memory_limit(start, size))) {
809 			if (--ranges)
810 				goto new_range;
811 			else
812 				continue;
813 		}
814 
815 		memblock_set_node(start, size, &memblock.memory, nid);
816 
817 		if (--ranges)
818 			goto new_range;
819 	}
820 
821 	/*
822 	 * Now do the same thing for each MEMBLOCK listed in the
823 	 * ibm,dynamic-memory property in the
824 	 * ibm,dynamic-reconfiguration-memory node.
825 	 */
826 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
827 	if (memory)
828 		parse_drconf_memory(memory);
829 
830 	return 0;
831 }
832 
833 static void __init setup_nonnuma(void)
834 {
835 	unsigned long top_of_ram = memblock_end_of_DRAM();
836 	unsigned long total_ram = memblock_phys_mem_size();
837 	unsigned long start_pfn, end_pfn;
838 	unsigned int nid = 0;
839 	struct memblock_region *reg;
840 
841 	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
842 	       top_of_ram, total_ram);
843 	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
844 	       (top_of_ram - total_ram) >> 20);
845 
846 	for_each_memblock(memory, reg) {
847 		start_pfn = memblock_region_memory_base_pfn(reg);
848 		end_pfn = memblock_region_memory_end_pfn(reg);
849 
850 		fake_numa_create_new_node(end_pfn, &nid);
851 		memblock_set_node(PFN_PHYS(start_pfn),
852 				  PFN_PHYS(end_pfn - start_pfn),
853 				  &memblock.memory, nid);
854 		node_set_online(nid);
855 	}
856 }
857 
858 void __init dump_numa_cpu_topology(void)
859 {
860 	unsigned int node;
861 	unsigned int cpu, count;
862 
863 	if (min_common_depth == -1 || !numa_enabled)
864 		return;
865 
866 	for_each_online_node(node) {
867 		printk(KERN_DEBUG "Node %d CPUs:", node);
868 
869 		count = 0;
870 		/*
871 		 * If we used a CPU iterator here we would miss printing
872 		 * the holes in the cpumap.
873 		 */
874 		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
875 			if (cpumask_test_cpu(cpu,
876 					node_to_cpumask_map[node])) {
877 				if (count == 0)
878 					printk(" %u", cpu);
879 				++count;
880 			} else {
881 				if (count > 1)
882 					printk("-%u", cpu - 1);
883 				count = 0;
884 			}
885 		}
886 
887 		if (count > 1)
888 			printk("-%u", nr_cpu_ids - 1);
889 		printk("\n");
890 	}
891 }
892 
893 static void __init dump_numa_memory_topology(void)
894 {
895 	unsigned int node;
896 	unsigned int count;
897 
898 	if (min_common_depth == -1 || !numa_enabled)
899 		return;
900 
901 	for_each_online_node(node) {
902 		unsigned long i;
903 
904 		printk(KERN_DEBUG "Node %d Memory:", node);
905 
906 		count = 0;
907 
908 		for (i = 0; i < memblock_end_of_DRAM();
909 		     i += (1 << SECTION_SIZE_BITS)) {
910 			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
911 				if (count == 0)
912 					printk(" 0x%lx", i);
913 				++count;
914 			} else {
915 				if (count > 0)
916 					printk("-0x%lx", i);
917 				count = 0;
918 			}
919 		}
920 
921 		if (count > 0)
922 			printk("-0x%lx", i);
923 		printk("\n");
924 	}
925 }
926 
927 /*
928  * Allocate some memory, satisfying the memblock or bootmem allocator where
929  * required. nid is the preferred node and end is the physical address of
930  * the highest address in the node.
931  *
932  * Returns the virtual address of the memory.
933  */
934 static void __init *careful_zallocation(int nid, unsigned long size,
935 				       unsigned long align,
936 				       unsigned long end_pfn)
937 {
938 	void *ret;
939 	int new_nid;
940 	unsigned long ret_paddr;
941 
942 	ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
943 
944 	/* retry over all memory */
945 	if (!ret_paddr)
946 		ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
947 
948 	if (!ret_paddr)
949 		panic("numa.c: cannot allocate %lu bytes for node %d",
950 		      size, nid);
951 
952 	ret = __va(ret_paddr);
953 
954 	/*
955 	 * We initialize the nodes in numeric order: 0, 1, 2...
956 	 * and hand over control from the MEMBLOCK allocator to the
957 	 * bootmem allocator.  If this function is called for
958 	 * node 5, then we know that all nodes <5 are using the
959 	 * bootmem allocator instead of the MEMBLOCK allocator.
960 	 *
961 	 * So, check the nid from which this allocation came
962 	 * and double check to see if we need to use bootmem
963 	 * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
964 	 * since it would be useless.
965 	 */
966 	new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
967 	if (new_nid < nid) {
968 		ret = __alloc_bootmem_node(NODE_DATA(new_nid),
969 				size, align, 0);
970 
971 		dbg("alloc_bootmem %p %lx\n", ret, size);
972 	}
973 
974 	memset(ret, 0, size);
975 	return ret;
976 }
977 
978 static struct notifier_block ppc64_numa_nb = {
979 	.notifier_call = cpu_numa_callback,
980 	.priority = 1 /* Must run before sched domains notifier. */
981 };
982 
983 static void __init mark_reserved_regions_for_nid(int nid)
984 {
985 	struct pglist_data *node = NODE_DATA(nid);
986 	struct memblock_region *reg;
987 
988 	for_each_memblock(reserved, reg) {
989 		unsigned long physbase = reg->base;
990 		unsigned long size = reg->size;
991 		unsigned long start_pfn = physbase >> PAGE_SHIFT;
992 		unsigned long end_pfn = PFN_UP(physbase + size);
993 		struct node_active_region node_ar;
994 		unsigned long node_end_pfn = pgdat_end_pfn(node);
995 
996 		/*
997 		 * Check to make sure that this memblock.reserved area is
998 		 * within the bounds of the node that we care about.
999 		 * Checking the nid of the start and end points is not
1000 		 * sufficient because the reserved area could span the
1001 		 * entire node.
1002 		 */
1003 		if (end_pfn <= node->node_start_pfn ||
1004 		    start_pfn >= node_end_pfn)
1005 			continue;
1006 
1007 		get_node_active_region(start_pfn, &node_ar);
1008 		while (start_pfn < end_pfn &&
1009 			node_ar.start_pfn < node_ar.end_pfn) {
1010 			unsigned long reserve_size = size;
1011 			/*
1012 			 * if reserved region extends past active region
1013 			 * then trim size to active region
1014 			 */
1015 			if (end_pfn > node_ar.end_pfn)
1016 				reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1017 					- physbase;
1018 			/*
1019 			 * Only worry about *this* node, others may not
1020 			 * yet have valid NODE_DATA().
1021 			 */
1022 			if (node_ar.nid == nid) {
1023 				dbg("reserve_bootmem %lx %lx nid=%d\n",
1024 					physbase, reserve_size, node_ar.nid);
1025 				reserve_bootmem_node(NODE_DATA(node_ar.nid),
1026 						physbase, reserve_size,
1027 						BOOTMEM_DEFAULT);
1028 			}
1029 			/*
1030 			 * if reserved region is contained in the active region
1031 			 * then done.
1032 			 */
1033 			if (end_pfn <= node_ar.end_pfn)
1034 				break;
1035 
1036 			/*
1037 			 * reserved region extends past the active region
1038 			 *   get next active region that contains this
1039 			 *   reserved region
1040 			 */
1041 			start_pfn = node_ar.end_pfn;
1042 			physbase = start_pfn << PAGE_SHIFT;
1043 			size = size - reserve_size;
1044 			get_node_active_region(start_pfn, &node_ar);
1045 		}
1046 	}
1047 }
1048 
1049 
1050 void __init do_init_bootmem(void)
1051 {
1052 	int nid;
1053 
1054 	min_low_pfn = 0;
1055 	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1056 	max_pfn = max_low_pfn;
1057 
1058 	if (parse_numa_properties())
1059 		setup_nonnuma();
1060 	else
1061 		dump_numa_memory_topology();
1062 
1063 	for_each_online_node(nid) {
1064 		unsigned long start_pfn, end_pfn;
1065 		void *bootmem_vaddr;
1066 		unsigned long bootmap_pages;
1067 
1068 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1069 
1070 		/*
1071 		 * Allocate the node structure node local if possible
1072 		 *
1073 		 * Be careful moving this around, as it relies on all
1074 		 * previous nodes' bootmem to be initialized and have
1075 		 * all reserved areas marked.
1076 		 */
1077 		NODE_DATA(nid) = careful_zallocation(nid,
1078 					sizeof(struct pglist_data),
1079 					SMP_CACHE_BYTES, end_pfn);
1080 
1081   		dbg("node %d\n", nid);
1082 		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1083 
1084 		NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1085 		NODE_DATA(nid)->node_start_pfn = start_pfn;
1086 		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1087 
1088 		if (NODE_DATA(nid)->node_spanned_pages == 0)
1089   			continue;
1090 
1091   		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1092   		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1093 
1094 		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1095 		bootmem_vaddr = careful_zallocation(nid,
1096 					bootmap_pages << PAGE_SHIFT,
1097 					PAGE_SIZE, end_pfn);
1098 
1099 		dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1100 
1101 		init_bootmem_node(NODE_DATA(nid),
1102 				  __pa(bootmem_vaddr) >> PAGE_SHIFT,
1103 				  start_pfn, end_pfn);
1104 
1105 		free_bootmem_with_active_regions(nid, end_pfn);
1106 		/*
1107 		 * Be very careful about moving this around.  Future
1108 		 * calls to careful_zallocation() depend on this getting
1109 		 * done correctly.
1110 		 */
1111 		mark_reserved_regions_for_nid(nid);
1112 		sparse_memory_present_with_active_regions(nid);
1113 	}
1114 
1115 	init_bootmem_done = 1;
1116 
1117 	/*
1118 	 * Now bootmem is initialised we can create the node to cpumask
1119 	 * lookup tables and setup the cpu callback to populate them.
1120 	 */
1121 	setup_node_to_cpumask_map();
1122 
1123 	reset_numa_cpu_lookup_table();
1124 	register_cpu_notifier(&ppc64_numa_nb);
1125 	cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1126 			  (void *)(unsigned long)boot_cpuid);
1127 }
1128 
1129 void __init paging_init(void)
1130 {
1131 	unsigned long max_zone_pfns[MAX_NR_ZONES];
1132 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1133 	max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1134 	free_area_init_nodes(max_zone_pfns);
1135 }
1136 
1137 static int __init early_numa(char *p)
1138 {
1139 	if (!p)
1140 		return 0;
1141 
1142 	if (strstr(p, "off"))
1143 		numa_enabled = 0;
1144 
1145 	if (strstr(p, "debug"))
1146 		numa_debug = 1;
1147 
1148 	p = strstr(p, "fake=");
1149 	if (p)
1150 		cmdline = p + strlen("fake=");
1151 
1152 	return 0;
1153 }
1154 early_param("numa", early_numa);
1155 
1156 #ifdef CONFIG_MEMORY_HOTPLUG
1157 /*
1158  * Find the node associated with a hot added memory section for
1159  * memory represented in the device tree by the property
1160  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1161  */
1162 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1163 				     unsigned long scn_addr)
1164 {
1165 	const __be32 *dm;
1166 	unsigned int drconf_cell_cnt, rc;
1167 	unsigned long lmb_size;
1168 	struct assoc_arrays aa;
1169 	int nid = -1;
1170 
1171 	drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1172 	if (!drconf_cell_cnt)
1173 		return -1;
1174 
1175 	lmb_size = of_get_lmb_size(memory);
1176 	if (!lmb_size)
1177 		return -1;
1178 
1179 	rc = of_get_assoc_arrays(memory, &aa);
1180 	if (rc)
1181 		return -1;
1182 
1183 	for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1184 		struct of_drconf_cell drmem;
1185 
1186 		read_drconf_cell(&drmem, &dm);
1187 
1188 		/* skip this block if it is reserved or not assigned to
1189 		 * this partition */
1190 		if ((drmem.flags & DRCONF_MEM_RESERVED)
1191 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1192 			continue;
1193 
1194 		if ((scn_addr < drmem.base_addr)
1195 		    || (scn_addr >= (drmem.base_addr + lmb_size)))
1196 			continue;
1197 
1198 		nid = of_drconf_to_nid_single(&drmem, &aa);
1199 		break;
1200 	}
1201 
1202 	return nid;
1203 }
1204 
1205 /*
1206  * Find the node associated with a hot added memory section for memory
1207  * represented in the device tree as a node (i.e. memory@XXXX) for
1208  * each memblock.
1209  */
1210 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1211 {
1212 	struct device_node *memory;
1213 	int nid = -1;
1214 
1215 	for_each_node_by_type(memory, "memory") {
1216 		unsigned long start, size;
1217 		int ranges;
1218 		const __be32 *memcell_buf;
1219 		unsigned int len;
1220 
1221 		memcell_buf = of_get_property(memory, "reg", &len);
1222 		if (!memcell_buf || len <= 0)
1223 			continue;
1224 
1225 		/* ranges in cell */
1226 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1227 
1228 		while (ranges--) {
1229 			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1230 			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1231 
1232 			if ((scn_addr < start) || (scn_addr >= (start + size)))
1233 				continue;
1234 
1235 			nid = of_node_to_nid_single(memory);
1236 			break;
1237 		}
1238 
1239 		if (nid >= 0)
1240 			break;
1241 	}
1242 
1243 	of_node_put(memory);
1244 
1245 	return nid;
1246 }
1247 
1248 /*
1249  * Find the node associated with a hot added memory section.  Section
1250  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1251  * sections are fully contained within a single MEMBLOCK.
1252  */
1253 int hot_add_scn_to_nid(unsigned long scn_addr)
1254 {
1255 	struct device_node *memory = NULL;
1256 	int nid, found = 0;
1257 
1258 	if (!numa_enabled || (min_common_depth < 0))
1259 		return first_online_node;
1260 
1261 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1262 	if (memory) {
1263 		nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1264 		of_node_put(memory);
1265 	} else {
1266 		nid = hot_add_node_scn_to_nid(scn_addr);
1267 	}
1268 
1269 	if (nid < 0 || !node_online(nid))
1270 		nid = first_online_node;
1271 
1272 	if (NODE_DATA(nid)->node_spanned_pages)
1273 		return nid;
1274 
1275 	for_each_online_node(nid) {
1276 		if (NODE_DATA(nid)->node_spanned_pages) {
1277 			found = 1;
1278 			break;
1279 		}
1280 	}
1281 
1282 	BUG_ON(!found);
1283 	return nid;
1284 }
1285 
1286 static u64 hot_add_drconf_memory_max(void)
1287 {
1288         struct device_node *memory = NULL;
1289         unsigned int drconf_cell_cnt = 0;
1290         u64 lmb_size = 0;
1291 	const __be32 *dm = NULL;
1292 
1293         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1294         if (memory) {
1295                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1296                 lmb_size = of_get_lmb_size(memory);
1297                 of_node_put(memory);
1298         }
1299         return lmb_size * drconf_cell_cnt;
1300 }
1301 
1302 /*
1303  * memory_hotplug_max - return max address of memory that may be added
1304  *
1305  * This is currently only used on systems that support drconfig memory
1306  * hotplug.
1307  */
1308 u64 memory_hotplug_max(void)
1309 {
1310         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1311 }
1312 #endif /* CONFIG_MEMORY_HOTPLUG */
1313 
1314 /* Virtual Processor Home Node (VPHN) support */
1315 #ifdef CONFIG_PPC_SPLPAR
1316 struct topology_update_data {
1317 	struct topology_update_data *next;
1318 	unsigned int cpu;
1319 	int old_nid;
1320 	int new_nid;
1321 };
1322 
1323 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1324 static cpumask_t cpu_associativity_changes_mask;
1325 static int vphn_enabled;
1326 static int prrn_enabled;
1327 static void reset_topology_timer(void);
1328 
1329 /*
1330  * Store the current values of the associativity change counters in the
1331  * hypervisor.
1332  */
1333 static void setup_cpu_associativity_change_counters(void)
1334 {
1335 	int cpu;
1336 
1337 	/* The VPHN feature supports a maximum of 8 reference points */
1338 	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1339 
1340 	for_each_possible_cpu(cpu) {
1341 		int i;
1342 		u8 *counts = vphn_cpu_change_counts[cpu];
1343 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1344 
1345 		for (i = 0; i < distance_ref_points_depth; i++)
1346 			counts[i] = hypervisor_counts[i];
1347 	}
1348 }
1349 
1350 /*
1351  * The hypervisor maintains a set of 8 associativity change counters in
1352  * the VPA of each cpu that correspond to the associativity levels in the
1353  * ibm,associativity-reference-points property. When an associativity
1354  * level changes, the corresponding counter is incremented.
1355  *
1356  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1357  * node associativity levels have changed.
1358  *
1359  * Returns the number of cpus with unhandled associativity changes.
1360  */
1361 static int update_cpu_associativity_changes_mask(void)
1362 {
1363 	int cpu;
1364 	cpumask_t *changes = &cpu_associativity_changes_mask;
1365 
1366 	for_each_possible_cpu(cpu) {
1367 		int i, changed = 0;
1368 		u8 *counts = vphn_cpu_change_counts[cpu];
1369 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1370 
1371 		for (i = 0; i < distance_ref_points_depth; i++) {
1372 			if (hypervisor_counts[i] != counts[i]) {
1373 				counts[i] = hypervisor_counts[i];
1374 				changed = 1;
1375 			}
1376 		}
1377 		if (changed) {
1378 			cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1379 			cpu = cpu_last_thread_sibling(cpu);
1380 		}
1381 	}
1382 
1383 	return cpumask_weight(changes);
1384 }
1385 
1386 /*
1387  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1388  * the complete property we have to add the length in the first cell.
1389  */
1390 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1391 
1392 /*
1393  * Convert the associativity domain numbers returned from the hypervisor
1394  * to the sequence they would appear in the ibm,associativity property.
1395  */
1396 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1397 {
1398 	int i, nr_assoc_doms = 0;
1399 	const __be16 *field = (const __be16 *) packed;
1400 
1401 #define VPHN_FIELD_UNUSED	(0xffff)
1402 #define VPHN_FIELD_MSB		(0x8000)
1403 #define VPHN_FIELD_MASK		(~VPHN_FIELD_MSB)
1404 
1405 	for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1406 		if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1407 			/* All significant fields processed, and remaining
1408 			 * fields contain the reserved value of all 1's.
1409 			 * Just store them.
1410 			 */
1411 			unpacked[i] = *((__be32 *)field);
1412 			field += 2;
1413 		} else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1414 			/* Data is in the lower 15 bits of this field */
1415 			unpacked[i] = cpu_to_be32(
1416 				be16_to_cpup(field) & VPHN_FIELD_MASK);
1417 			field++;
1418 			nr_assoc_doms++;
1419 		} else {
1420 			/* Data is in the lower 15 bits of this field
1421 			 * concatenated with the next 16 bit field
1422 			 */
1423 			unpacked[i] = *((__be32 *)field);
1424 			field += 2;
1425 			nr_assoc_doms++;
1426 		}
1427 	}
1428 
1429 	/* The first cell contains the length of the property */
1430 	unpacked[0] = cpu_to_be32(nr_assoc_doms);
1431 
1432 	return nr_assoc_doms;
1433 }
1434 
1435 /*
1436  * Retrieve the new associativity information for a virtual processor's
1437  * home node.
1438  */
1439 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1440 {
1441 	long rc;
1442 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1443 	u64 flags = 1;
1444 	int hwcpu = get_hard_smp_processor_id(cpu);
1445 
1446 	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1447 	vphn_unpack_associativity(retbuf, associativity);
1448 
1449 	return rc;
1450 }
1451 
1452 static long vphn_get_associativity(unsigned long cpu,
1453 					__be32 *associativity)
1454 {
1455 	long rc;
1456 
1457 	rc = hcall_vphn(cpu, associativity);
1458 
1459 	switch (rc) {
1460 	case H_FUNCTION:
1461 		printk(KERN_INFO
1462 			"VPHN is not supported. Disabling polling...\n");
1463 		stop_topology_update();
1464 		break;
1465 	case H_HARDWARE:
1466 		printk(KERN_ERR
1467 			"hcall_vphn() experienced a hardware fault "
1468 			"preventing VPHN. Disabling polling...\n");
1469 		stop_topology_update();
1470 	}
1471 
1472 	return rc;
1473 }
1474 
1475 /*
1476  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1477  * characteristics change. This function doesn't perform any locking and is
1478  * only safe to call from stop_machine().
1479  */
1480 static int update_cpu_topology(void *data)
1481 {
1482 	struct topology_update_data *update;
1483 	unsigned long cpu;
1484 
1485 	if (!data)
1486 		return -EINVAL;
1487 
1488 	cpu = smp_processor_id();
1489 
1490 	for (update = data; update; update = update->next) {
1491 		if (cpu != update->cpu)
1492 			continue;
1493 
1494 		unmap_cpu_from_node(update->cpu);
1495 		map_cpu_to_node(update->cpu, update->new_nid);
1496 		vdso_getcpu_init();
1497 	}
1498 
1499 	return 0;
1500 }
1501 
1502 static int update_lookup_table(void *data)
1503 {
1504 	struct topology_update_data *update;
1505 
1506 	if (!data)
1507 		return -EINVAL;
1508 
1509 	/*
1510 	 * Upon topology update, the numa-cpu lookup table needs to be updated
1511 	 * for all threads in the core, including offline CPUs, to ensure that
1512 	 * future hotplug operations respect the cpu-to-node associativity
1513 	 * properly.
1514 	 */
1515 	for (update = data; update; update = update->next) {
1516 		int nid, base, j;
1517 
1518 		nid = update->new_nid;
1519 		base = cpu_first_thread_sibling(update->cpu);
1520 
1521 		for (j = 0; j < threads_per_core; j++) {
1522 			update_numa_cpu_lookup_table(base + j, nid);
1523 		}
1524 	}
1525 
1526 	return 0;
1527 }
1528 
1529 /*
1530  * Update the node maps and sysfs entries for each cpu whose home node
1531  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1532  */
1533 int arch_update_cpu_topology(void)
1534 {
1535 	unsigned int cpu, sibling, changed = 0;
1536 	struct topology_update_data *updates, *ud;
1537 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1538 	cpumask_t updated_cpus;
1539 	struct device *dev;
1540 	int weight, new_nid, i = 0;
1541 
1542 	weight = cpumask_weight(&cpu_associativity_changes_mask);
1543 	if (!weight)
1544 		return 0;
1545 
1546 	updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1547 	if (!updates)
1548 		return 0;
1549 
1550 	cpumask_clear(&updated_cpus);
1551 
1552 	for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1553 		/*
1554 		 * If siblings aren't flagged for changes, updates list
1555 		 * will be too short. Skip on this update and set for next
1556 		 * update.
1557 		 */
1558 		if (!cpumask_subset(cpu_sibling_mask(cpu),
1559 					&cpu_associativity_changes_mask)) {
1560 			pr_info("Sibling bits not set for associativity "
1561 					"change, cpu%d\n", cpu);
1562 			cpumask_or(&cpu_associativity_changes_mask,
1563 					&cpu_associativity_changes_mask,
1564 					cpu_sibling_mask(cpu));
1565 			cpu = cpu_last_thread_sibling(cpu);
1566 			continue;
1567 		}
1568 
1569 		/* Use associativity from first thread for all siblings */
1570 		vphn_get_associativity(cpu, associativity);
1571 		new_nid = associativity_to_nid(associativity);
1572 		if (new_nid < 0 || !node_online(new_nid))
1573 			new_nid = first_online_node;
1574 
1575 		if (new_nid == numa_cpu_lookup_table[cpu]) {
1576 			cpumask_andnot(&cpu_associativity_changes_mask,
1577 					&cpu_associativity_changes_mask,
1578 					cpu_sibling_mask(cpu));
1579 			cpu = cpu_last_thread_sibling(cpu);
1580 			continue;
1581 		}
1582 
1583 		for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1584 			ud = &updates[i++];
1585 			ud->cpu = sibling;
1586 			ud->new_nid = new_nid;
1587 			ud->old_nid = numa_cpu_lookup_table[sibling];
1588 			cpumask_set_cpu(sibling, &updated_cpus);
1589 			if (i < weight)
1590 				ud->next = &updates[i];
1591 		}
1592 		cpu = cpu_last_thread_sibling(cpu);
1593 	}
1594 
1595 	/*
1596 	 * In cases where we have nothing to update (because the updates list
1597 	 * is too short or because the new topology is same as the old one),
1598 	 * skip invoking update_cpu_topology() via stop-machine(). This is
1599 	 * necessary (and not just a fast-path optimization) since stop-machine
1600 	 * can end up electing a random CPU to run update_cpu_topology(), and
1601 	 * thus trick us into setting up incorrect cpu-node mappings (since
1602 	 * 'updates' is kzalloc()'ed).
1603 	 *
1604 	 * And for the similar reason, we will skip all the following updating.
1605 	 */
1606 	if (!cpumask_weight(&updated_cpus))
1607 		goto out;
1608 
1609 	stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1610 
1611 	/*
1612 	 * Update the numa-cpu lookup table with the new mappings, even for
1613 	 * offline CPUs. It is best to perform this update from the stop-
1614 	 * machine context.
1615 	 */
1616 	stop_machine(update_lookup_table, &updates[0],
1617 					cpumask_of(raw_smp_processor_id()));
1618 
1619 	for (ud = &updates[0]; ud; ud = ud->next) {
1620 		unregister_cpu_under_node(ud->cpu, ud->old_nid);
1621 		register_cpu_under_node(ud->cpu, ud->new_nid);
1622 
1623 		dev = get_cpu_device(ud->cpu);
1624 		if (dev)
1625 			kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1626 		cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1627 		changed = 1;
1628 	}
1629 
1630 out:
1631 	kfree(updates);
1632 	return changed;
1633 }
1634 
1635 static void topology_work_fn(struct work_struct *work)
1636 {
1637 	rebuild_sched_domains();
1638 }
1639 static DECLARE_WORK(topology_work, topology_work_fn);
1640 
1641 static void topology_schedule_update(void)
1642 {
1643 	schedule_work(&topology_work);
1644 }
1645 
1646 static void topology_timer_fn(unsigned long ignored)
1647 {
1648 	if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1649 		topology_schedule_update();
1650 	else if (vphn_enabled) {
1651 		if (update_cpu_associativity_changes_mask() > 0)
1652 			topology_schedule_update();
1653 		reset_topology_timer();
1654 	}
1655 }
1656 static struct timer_list topology_timer =
1657 	TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1658 
1659 static void reset_topology_timer(void)
1660 {
1661 	topology_timer.data = 0;
1662 	topology_timer.expires = jiffies + 60 * HZ;
1663 	mod_timer(&topology_timer, topology_timer.expires);
1664 }
1665 
1666 #ifdef CONFIG_SMP
1667 
1668 static void stage_topology_update(int core_id)
1669 {
1670 	cpumask_or(&cpu_associativity_changes_mask,
1671 		&cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1672 	reset_topology_timer();
1673 }
1674 
1675 static int dt_update_callback(struct notifier_block *nb,
1676 				unsigned long action, void *data)
1677 {
1678 	struct of_prop_reconfig *update;
1679 	int rc = NOTIFY_DONE;
1680 
1681 	switch (action) {
1682 	case OF_RECONFIG_UPDATE_PROPERTY:
1683 		update = (struct of_prop_reconfig *)data;
1684 		if (!of_prop_cmp(update->dn->type, "cpu") &&
1685 		    !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1686 			u32 core_id;
1687 			of_property_read_u32(update->dn, "reg", &core_id);
1688 			stage_topology_update(core_id);
1689 			rc = NOTIFY_OK;
1690 		}
1691 		break;
1692 	}
1693 
1694 	return rc;
1695 }
1696 
1697 static struct notifier_block dt_update_nb = {
1698 	.notifier_call = dt_update_callback,
1699 };
1700 
1701 #endif
1702 
1703 /*
1704  * Start polling for associativity changes.
1705  */
1706 int start_topology_update(void)
1707 {
1708 	int rc = 0;
1709 
1710 	if (firmware_has_feature(FW_FEATURE_PRRN)) {
1711 		if (!prrn_enabled) {
1712 			prrn_enabled = 1;
1713 			vphn_enabled = 0;
1714 #ifdef CONFIG_SMP
1715 			rc = of_reconfig_notifier_register(&dt_update_nb);
1716 #endif
1717 		}
1718 	} else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1719 		   lppaca_shared_proc(get_lppaca())) {
1720 		if (!vphn_enabled) {
1721 			prrn_enabled = 0;
1722 			vphn_enabled = 1;
1723 			setup_cpu_associativity_change_counters();
1724 			init_timer_deferrable(&topology_timer);
1725 			reset_topology_timer();
1726 		}
1727 	}
1728 
1729 	return rc;
1730 }
1731 
1732 /*
1733  * Disable polling for VPHN associativity changes.
1734  */
1735 int stop_topology_update(void)
1736 {
1737 	int rc = 0;
1738 
1739 	if (prrn_enabled) {
1740 		prrn_enabled = 0;
1741 #ifdef CONFIG_SMP
1742 		rc = of_reconfig_notifier_unregister(&dt_update_nb);
1743 #endif
1744 	} else if (vphn_enabled) {
1745 		vphn_enabled = 0;
1746 		rc = del_timer_sync(&topology_timer);
1747 	}
1748 
1749 	return rc;
1750 }
1751 
1752 int prrn_is_enabled(void)
1753 {
1754 	return prrn_enabled;
1755 }
1756 
1757 static int topology_read(struct seq_file *file, void *v)
1758 {
1759 	if (vphn_enabled || prrn_enabled)
1760 		seq_puts(file, "on\n");
1761 	else
1762 		seq_puts(file, "off\n");
1763 
1764 	return 0;
1765 }
1766 
1767 static int topology_open(struct inode *inode, struct file *file)
1768 {
1769 	return single_open(file, topology_read, NULL);
1770 }
1771 
1772 static ssize_t topology_write(struct file *file, const char __user *buf,
1773 			      size_t count, loff_t *off)
1774 {
1775 	char kbuf[4]; /* "on" or "off" plus null. */
1776 	int read_len;
1777 
1778 	read_len = count < 3 ? count : 3;
1779 	if (copy_from_user(kbuf, buf, read_len))
1780 		return -EINVAL;
1781 
1782 	kbuf[read_len] = '\0';
1783 
1784 	if (!strncmp(kbuf, "on", 2))
1785 		start_topology_update();
1786 	else if (!strncmp(kbuf, "off", 3))
1787 		stop_topology_update();
1788 	else
1789 		return -EINVAL;
1790 
1791 	return count;
1792 }
1793 
1794 static const struct file_operations topology_ops = {
1795 	.read = seq_read,
1796 	.write = topology_write,
1797 	.open = topology_open,
1798 	.release = single_release
1799 };
1800 
1801 static int topology_update_init(void)
1802 {
1803 	start_topology_update();
1804 	proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops);
1805 
1806 	return 0;
1807 }
1808 device_initcall(topology_update_init);
1809 #endif /* CONFIG_PPC_SPLPAR */
1810