xref: /openbmc/linux/arch/powerpc/mm/numa.c (revision 4a44a19b)
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 #define pr_fmt(fmt) "numa: " fmt
12 
13 #include <linux/threads.h>
14 #include <linux/bootmem.h>
15 #include <linux/init.h>
16 #include <linux/mm.h>
17 #include <linux/mmzone.h>
18 #include <linux/export.h>
19 #include <linux/nodemask.h>
20 #include <linux/cpu.h>
21 #include <linux/notifier.h>
22 #include <linux/memblock.h>
23 #include <linux/of.h>
24 #include <linux/pfn.h>
25 #include <linux/cpuset.h>
26 #include <linux/node.h>
27 #include <linux/stop_machine.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/uaccess.h>
31 #include <linux/slab.h>
32 #include <asm/cputhreads.h>
33 #include <asm/sparsemem.h>
34 #include <asm/prom.h>
35 #include <asm/smp.h>
36 #include <asm/cputhreads.h>
37 #include <asm/topology.h>
38 #include <asm/firmware.h>
39 #include <asm/paca.h>
40 #include <asm/hvcall.h>
41 #include <asm/setup.h>
42 #include <asm/vdso.h>
43 
44 static int numa_enabled = 1;
45 
46 static char *cmdline __initdata;
47 
48 static int numa_debug;
49 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
50 
51 int numa_cpu_lookup_table[NR_CPUS];
52 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
53 struct pglist_data *node_data[MAX_NUMNODES];
54 
55 EXPORT_SYMBOL(numa_cpu_lookup_table);
56 EXPORT_SYMBOL(node_to_cpumask_map);
57 EXPORT_SYMBOL(node_data);
58 
59 static int min_common_depth;
60 static int n_mem_addr_cells, n_mem_size_cells;
61 static int form1_affinity;
62 
63 #define MAX_DISTANCE_REF_POINTS 4
64 static int distance_ref_points_depth;
65 static const __be32 *distance_ref_points;
66 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
67 
68 /*
69  * Allocate node_to_cpumask_map based on number of available nodes
70  * Requires node_possible_map to be valid.
71  *
72  * Note: cpumask_of_node() is not valid until after this is done.
73  */
74 static void __init setup_node_to_cpumask_map(void)
75 {
76 	unsigned int node;
77 
78 	/* setup nr_node_ids if not done yet */
79 	if (nr_node_ids == MAX_NUMNODES)
80 		setup_nr_node_ids();
81 
82 	/* allocate the map */
83 	for (node = 0; node < nr_node_ids; node++)
84 		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
85 
86 	/* cpumask_of_node() will now work */
87 	dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
88 }
89 
90 static int __init fake_numa_create_new_node(unsigned long end_pfn,
91 						unsigned int *nid)
92 {
93 	unsigned long long mem;
94 	char *p = cmdline;
95 	static unsigned int fake_nid;
96 	static unsigned long long curr_boundary;
97 
98 	/*
99 	 * Modify node id, iff we started creating NUMA nodes
100 	 * We want to continue from where we left of the last time
101 	 */
102 	if (fake_nid)
103 		*nid = fake_nid;
104 	/*
105 	 * In case there are no more arguments to parse, the
106 	 * node_id should be the same as the last fake node id
107 	 * (we've handled this above).
108 	 */
109 	if (!p)
110 		return 0;
111 
112 	mem = memparse(p, &p);
113 	if (!mem)
114 		return 0;
115 
116 	if (mem < curr_boundary)
117 		return 0;
118 
119 	curr_boundary = mem;
120 
121 	if ((end_pfn << PAGE_SHIFT) > mem) {
122 		/*
123 		 * Skip commas and spaces
124 		 */
125 		while (*p == ',' || *p == ' ' || *p == '\t')
126 			p++;
127 
128 		cmdline = p;
129 		fake_nid++;
130 		*nid = fake_nid;
131 		dbg("created new fake_node with id %d\n", fake_nid);
132 		return 1;
133 	}
134 	return 0;
135 }
136 
137 /*
138  * get_node_active_region - Return active region containing pfn
139  * Active range returned is empty if none found.
140  * @pfn: The page to return the region for
141  * @node_ar: Returned set to the active region containing @pfn
142  */
143 static void __init get_node_active_region(unsigned long pfn,
144 					  struct node_active_region *node_ar)
145 {
146 	unsigned long start_pfn, end_pfn;
147 	int i, nid;
148 
149 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
150 		if (pfn >= start_pfn && pfn < end_pfn) {
151 			node_ar->nid = nid;
152 			node_ar->start_pfn = start_pfn;
153 			node_ar->end_pfn = end_pfn;
154 			break;
155 		}
156 	}
157 }
158 
159 static void reset_numa_cpu_lookup_table(void)
160 {
161 	unsigned int cpu;
162 
163 	for_each_possible_cpu(cpu)
164 		numa_cpu_lookup_table[cpu] = -1;
165 }
166 
167 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
168 {
169 	numa_cpu_lookup_table[cpu] = node;
170 }
171 
172 static void map_cpu_to_node(int cpu, int node)
173 {
174 	update_numa_cpu_lookup_table(cpu, node);
175 
176 	dbg("adding cpu %d to node %d\n", cpu, node);
177 
178 	if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
179 		cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
180 }
181 
182 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
183 static void unmap_cpu_from_node(unsigned long cpu)
184 {
185 	int node = numa_cpu_lookup_table[cpu];
186 
187 	dbg("removing cpu %lu from node %d\n", cpu, node);
188 
189 	if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
190 		cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
191 	} else {
192 		printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
193 		       cpu, node);
194 	}
195 }
196 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
197 
198 /* must hold reference to node during call */
199 static const __be32 *of_get_associativity(struct device_node *dev)
200 {
201 	return of_get_property(dev, "ibm,associativity", NULL);
202 }
203 
204 /*
205  * Returns the property linux,drconf-usable-memory if
206  * it exists (the property exists only in kexec/kdump kernels,
207  * added by kexec-tools)
208  */
209 static const __be32 *of_get_usable_memory(struct device_node *memory)
210 {
211 	const __be32 *prop;
212 	u32 len;
213 	prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
214 	if (!prop || len < sizeof(unsigned int))
215 		return NULL;
216 	return prop;
217 }
218 
219 int __node_distance(int a, int b)
220 {
221 	int i;
222 	int distance = LOCAL_DISTANCE;
223 
224 	if (!form1_affinity)
225 		return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
226 
227 	for (i = 0; i < distance_ref_points_depth; i++) {
228 		if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
229 			break;
230 
231 		/* Double the distance for each NUMA level */
232 		distance *= 2;
233 	}
234 
235 	return distance;
236 }
237 EXPORT_SYMBOL(__node_distance);
238 
239 static void initialize_distance_lookup_table(int nid,
240 		const __be32 *associativity)
241 {
242 	int i;
243 
244 	if (!form1_affinity)
245 		return;
246 
247 	for (i = 0; i < distance_ref_points_depth; i++) {
248 		const __be32 *entry;
249 
250 		entry = &associativity[be32_to_cpu(distance_ref_points[i])];
251 		distance_lookup_table[nid][i] = of_read_number(entry, 1);
252 	}
253 }
254 
255 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
256  * info is found.
257  */
258 static int associativity_to_nid(const __be32 *associativity)
259 {
260 	int nid = -1;
261 
262 	if (min_common_depth == -1)
263 		goto out;
264 
265 	if (of_read_number(associativity, 1) >= min_common_depth)
266 		nid = of_read_number(&associativity[min_common_depth], 1);
267 
268 	/* POWER4 LPAR uses 0xffff as invalid node */
269 	if (nid == 0xffff || nid >= MAX_NUMNODES)
270 		nid = -1;
271 
272 	if (nid > 0 &&
273 	    of_read_number(associativity, 1) >= distance_ref_points_depth)
274 		initialize_distance_lookup_table(nid, associativity);
275 
276 out:
277 	return nid;
278 }
279 
280 /* Returns the nid associated with the given device tree node,
281  * or -1 if not found.
282  */
283 static int of_node_to_nid_single(struct device_node *device)
284 {
285 	int nid = -1;
286 	const __be32 *tmp;
287 
288 	tmp = of_get_associativity(device);
289 	if (tmp)
290 		nid = associativity_to_nid(tmp);
291 	return nid;
292 }
293 
294 /* Walk the device tree upwards, looking for an associativity id */
295 int of_node_to_nid(struct device_node *device)
296 {
297 	struct device_node *tmp;
298 	int nid = -1;
299 
300 	of_node_get(device);
301 	while (device) {
302 		nid = of_node_to_nid_single(device);
303 		if (nid != -1)
304 			break;
305 
306 	        tmp = device;
307 		device = of_get_parent(tmp);
308 		of_node_put(tmp);
309 	}
310 	of_node_put(device);
311 
312 	return nid;
313 }
314 EXPORT_SYMBOL_GPL(of_node_to_nid);
315 
316 static int __init find_min_common_depth(void)
317 {
318 	int depth;
319 	struct device_node *root;
320 
321 	if (firmware_has_feature(FW_FEATURE_OPAL))
322 		root = of_find_node_by_path("/ibm,opal");
323 	else
324 		root = of_find_node_by_path("/rtas");
325 	if (!root)
326 		root = of_find_node_by_path("/");
327 
328 	/*
329 	 * This property is a set of 32-bit integers, each representing
330 	 * an index into the ibm,associativity nodes.
331 	 *
332 	 * With form 0 affinity the first integer is for an SMP configuration
333 	 * (should be all 0's) and the second is for a normal NUMA
334 	 * configuration. We have only one level of NUMA.
335 	 *
336 	 * With form 1 affinity the first integer is the most significant
337 	 * NUMA boundary and the following are progressively less significant
338 	 * boundaries. There can be more than one level of NUMA.
339 	 */
340 	distance_ref_points = of_get_property(root,
341 					"ibm,associativity-reference-points",
342 					&distance_ref_points_depth);
343 
344 	if (!distance_ref_points) {
345 		dbg("NUMA: ibm,associativity-reference-points not found.\n");
346 		goto err;
347 	}
348 
349 	distance_ref_points_depth /= sizeof(int);
350 
351 	if (firmware_has_feature(FW_FEATURE_OPAL) ||
352 	    firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
353 		dbg("Using form 1 affinity\n");
354 		form1_affinity = 1;
355 	}
356 
357 	if (form1_affinity) {
358 		depth = of_read_number(distance_ref_points, 1);
359 	} else {
360 		if (distance_ref_points_depth < 2) {
361 			printk(KERN_WARNING "NUMA: "
362 				"short ibm,associativity-reference-points\n");
363 			goto err;
364 		}
365 
366 		depth = of_read_number(&distance_ref_points[1], 1);
367 	}
368 
369 	/*
370 	 * Warn and cap if the hardware supports more than
371 	 * MAX_DISTANCE_REF_POINTS domains.
372 	 */
373 	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
374 		printk(KERN_WARNING "NUMA: distance array capped at "
375 			"%d entries\n", MAX_DISTANCE_REF_POINTS);
376 		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
377 	}
378 
379 	of_node_put(root);
380 	return depth;
381 
382 err:
383 	of_node_put(root);
384 	return -1;
385 }
386 
387 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
388 {
389 	struct device_node *memory = NULL;
390 
391 	memory = of_find_node_by_type(memory, "memory");
392 	if (!memory)
393 		panic("numa.c: No memory nodes found!");
394 
395 	*n_addr_cells = of_n_addr_cells(memory);
396 	*n_size_cells = of_n_size_cells(memory);
397 	of_node_put(memory);
398 }
399 
400 static unsigned long read_n_cells(int n, const __be32 **buf)
401 {
402 	unsigned long result = 0;
403 
404 	while (n--) {
405 		result = (result << 32) | of_read_number(*buf, 1);
406 		(*buf)++;
407 	}
408 	return result;
409 }
410 
411 /*
412  * Read the next memblock list entry from the ibm,dynamic-memory property
413  * and return the information in the provided of_drconf_cell structure.
414  */
415 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
416 {
417 	const __be32 *cp;
418 
419 	drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
420 
421 	cp = *cellp;
422 	drmem->drc_index = of_read_number(cp, 1);
423 	drmem->reserved = of_read_number(&cp[1], 1);
424 	drmem->aa_index = of_read_number(&cp[2], 1);
425 	drmem->flags = of_read_number(&cp[3], 1);
426 
427 	*cellp = cp + 4;
428 }
429 
430 /*
431  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
432  *
433  * The layout of the ibm,dynamic-memory property is a number N of memblock
434  * list entries followed by N memblock list entries.  Each memblock list entry
435  * contains information as laid out in the of_drconf_cell struct above.
436  */
437 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
438 {
439 	const __be32 *prop;
440 	u32 len, entries;
441 
442 	prop = of_get_property(memory, "ibm,dynamic-memory", &len);
443 	if (!prop || len < sizeof(unsigned int))
444 		return 0;
445 
446 	entries = of_read_number(prop++, 1);
447 
448 	/* Now that we know the number of entries, revalidate the size
449 	 * of the property read in to ensure we have everything
450 	 */
451 	if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
452 		return 0;
453 
454 	*dm = prop;
455 	return entries;
456 }
457 
458 /*
459  * Retrieve and validate the ibm,lmb-size property for drconf memory
460  * from the device tree.
461  */
462 static u64 of_get_lmb_size(struct device_node *memory)
463 {
464 	const __be32 *prop;
465 	u32 len;
466 
467 	prop = of_get_property(memory, "ibm,lmb-size", &len);
468 	if (!prop || len < sizeof(unsigned int))
469 		return 0;
470 
471 	return read_n_cells(n_mem_size_cells, &prop);
472 }
473 
474 struct assoc_arrays {
475 	u32	n_arrays;
476 	u32	array_sz;
477 	const __be32 *arrays;
478 };
479 
480 /*
481  * Retrieve and validate the list of associativity arrays for drconf
482  * memory from the ibm,associativity-lookup-arrays property of the
483  * device tree..
484  *
485  * The layout of the ibm,associativity-lookup-arrays property is a number N
486  * indicating the number of associativity arrays, followed by a number M
487  * indicating the size of each associativity array, followed by a list
488  * of N associativity arrays.
489  */
490 static int of_get_assoc_arrays(struct device_node *memory,
491 			       struct assoc_arrays *aa)
492 {
493 	const __be32 *prop;
494 	u32 len;
495 
496 	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
497 	if (!prop || len < 2 * sizeof(unsigned int))
498 		return -1;
499 
500 	aa->n_arrays = of_read_number(prop++, 1);
501 	aa->array_sz = of_read_number(prop++, 1);
502 
503 	/* Now that we know the number of arrays and size of each array,
504 	 * revalidate the size of the property read in.
505 	 */
506 	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
507 		return -1;
508 
509 	aa->arrays = prop;
510 	return 0;
511 }
512 
513 /*
514  * This is like of_node_to_nid_single() for memory represented in the
515  * ibm,dynamic-reconfiguration-memory node.
516  */
517 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
518 				   struct assoc_arrays *aa)
519 {
520 	int default_nid = 0;
521 	int nid = default_nid;
522 	int index;
523 
524 	if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
525 	    !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
526 	    drmem->aa_index < aa->n_arrays) {
527 		index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
528 		nid = of_read_number(&aa->arrays[index], 1);
529 
530 		if (nid == 0xffff || nid >= MAX_NUMNODES)
531 			nid = default_nid;
532 	}
533 
534 	return nid;
535 }
536 
537 /*
538  * Figure out to which domain a cpu belongs and stick it there.
539  * Return the id of the domain used.
540  */
541 static int numa_setup_cpu(unsigned long lcpu)
542 {
543 	int nid = -1;
544 	struct device_node *cpu;
545 
546 	/*
547 	 * If a valid cpu-to-node mapping is already available, use it
548 	 * directly instead of querying the firmware, since it represents
549 	 * the most recent mapping notified to us by the platform (eg: VPHN).
550 	 */
551 	if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
552 		map_cpu_to_node(lcpu, nid);
553 		return nid;
554 	}
555 
556 	cpu = of_get_cpu_node(lcpu, NULL);
557 
558 	if (!cpu) {
559 		WARN_ON(1);
560 		if (cpu_present(lcpu))
561 			goto out_present;
562 		else
563 			goto out;
564 	}
565 
566 	nid = of_node_to_nid_single(cpu);
567 
568 out_present:
569 	if (nid < 0 || !node_online(nid))
570 		nid = first_online_node;
571 
572 	map_cpu_to_node(lcpu, nid);
573 	of_node_put(cpu);
574 out:
575 	return nid;
576 }
577 
578 static void verify_cpu_node_mapping(int cpu, int node)
579 {
580 	int base, sibling, i;
581 
582 	/* Verify that all the threads in the core belong to the same node */
583 	base = cpu_first_thread_sibling(cpu);
584 
585 	for (i = 0; i < threads_per_core; i++) {
586 		sibling = base + i;
587 
588 		if (sibling == cpu || cpu_is_offline(sibling))
589 			continue;
590 
591 		if (cpu_to_node(sibling) != node) {
592 			WARN(1, "CPU thread siblings %d and %d don't belong"
593 				" to the same node!\n", cpu, sibling);
594 			break;
595 		}
596 	}
597 }
598 
599 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
600 			     void *hcpu)
601 {
602 	unsigned long lcpu = (unsigned long)hcpu;
603 	int ret = NOTIFY_DONE, nid;
604 
605 	switch (action) {
606 	case CPU_UP_PREPARE:
607 	case CPU_UP_PREPARE_FROZEN:
608 		nid = numa_setup_cpu(lcpu);
609 		verify_cpu_node_mapping((int)lcpu, nid);
610 		ret = NOTIFY_OK;
611 		break;
612 #ifdef CONFIG_HOTPLUG_CPU
613 	case CPU_DEAD:
614 	case CPU_DEAD_FROZEN:
615 	case CPU_UP_CANCELED:
616 	case CPU_UP_CANCELED_FROZEN:
617 		unmap_cpu_from_node(lcpu);
618 		ret = NOTIFY_OK;
619 		break;
620 #endif
621 	}
622 	return ret;
623 }
624 
625 /*
626  * Check and possibly modify a memory region to enforce the memory limit.
627  *
628  * Returns the size the region should have to enforce the memory limit.
629  * This will either be the original value of size, a truncated value,
630  * or zero. If the returned value of size is 0 the region should be
631  * discarded as it lies wholly above the memory limit.
632  */
633 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
634 						      unsigned long size)
635 {
636 	/*
637 	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
638 	 * we've already adjusted it for the limit and it takes care of
639 	 * having memory holes below the limit.  Also, in the case of
640 	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
641 	 */
642 
643 	if (start + size <= memblock_end_of_DRAM())
644 		return size;
645 
646 	if (start >= memblock_end_of_DRAM())
647 		return 0;
648 
649 	return memblock_end_of_DRAM() - start;
650 }
651 
652 /*
653  * Reads the counter for a given entry in
654  * linux,drconf-usable-memory property
655  */
656 static inline int __init read_usm_ranges(const __be32 **usm)
657 {
658 	/*
659 	 * For each lmb in ibm,dynamic-memory a corresponding
660 	 * entry in linux,drconf-usable-memory property contains
661 	 * a counter followed by that many (base, size) duple.
662 	 * read the counter from linux,drconf-usable-memory
663 	 */
664 	return read_n_cells(n_mem_size_cells, usm);
665 }
666 
667 /*
668  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
669  * node.  This assumes n_mem_{addr,size}_cells have been set.
670  */
671 static void __init parse_drconf_memory(struct device_node *memory)
672 {
673 	const __be32 *uninitialized_var(dm), *usm;
674 	unsigned int n, rc, ranges, is_kexec_kdump = 0;
675 	unsigned long lmb_size, base, size, sz;
676 	int nid;
677 	struct assoc_arrays aa = { .arrays = NULL };
678 
679 	n = of_get_drconf_memory(memory, &dm);
680 	if (!n)
681 		return;
682 
683 	lmb_size = of_get_lmb_size(memory);
684 	if (!lmb_size)
685 		return;
686 
687 	rc = of_get_assoc_arrays(memory, &aa);
688 	if (rc)
689 		return;
690 
691 	/* check if this is a kexec/kdump kernel */
692 	usm = of_get_usable_memory(memory);
693 	if (usm != NULL)
694 		is_kexec_kdump = 1;
695 
696 	for (; n != 0; --n) {
697 		struct of_drconf_cell drmem;
698 
699 		read_drconf_cell(&drmem, &dm);
700 
701 		/* skip this block if the reserved bit is set in flags (0x80)
702 		   or if the block is not assigned to this partition (0x8) */
703 		if ((drmem.flags & DRCONF_MEM_RESERVED)
704 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
705 			continue;
706 
707 		base = drmem.base_addr;
708 		size = lmb_size;
709 		ranges = 1;
710 
711 		if (is_kexec_kdump) {
712 			ranges = read_usm_ranges(&usm);
713 			if (!ranges) /* there are no (base, size) duple */
714 				continue;
715 		}
716 		do {
717 			if (is_kexec_kdump) {
718 				base = read_n_cells(n_mem_addr_cells, &usm);
719 				size = read_n_cells(n_mem_size_cells, &usm);
720 			}
721 			nid = of_drconf_to_nid_single(&drmem, &aa);
722 			fake_numa_create_new_node(
723 				((base + size) >> PAGE_SHIFT),
724 					   &nid);
725 			node_set_online(nid);
726 			sz = numa_enforce_memory_limit(base, size);
727 			if (sz)
728 				memblock_set_node(base, sz,
729 						  &memblock.memory, nid);
730 		} while (--ranges);
731 	}
732 }
733 
734 static int __init parse_numa_properties(void)
735 {
736 	struct device_node *memory;
737 	int default_nid = 0;
738 	unsigned long i;
739 
740 	if (numa_enabled == 0) {
741 		printk(KERN_WARNING "NUMA disabled by user\n");
742 		return -1;
743 	}
744 
745 	min_common_depth = find_min_common_depth();
746 
747 	if (min_common_depth < 0)
748 		return min_common_depth;
749 
750 	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
751 
752 	/*
753 	 * Even though we connect cpus to numa domains later in SMP
754 	 * init, we need to know the node ids now. This is because
755 	 * each node to be onlined must have NODE_DATA etc backing it.
756 	 */
757 	for_each_present_cpu(i) {
758 		struct device_node *cpu;
759 		int nid;
760 
761 		cpu = of_get_cpu_node(i, NULL);
762 		BUG_ON(!cpu);
763 		nid = of_node_to_nid_single(cpu);
764 		of_node_put(cpu);
765 
766 		/*
767 		 * Don't fall back to default_nid yet -- we will plug
768 		 * cpus into nodes once the memory scan has discovered
769 		 * the topology.
770 		 */
771 		if (nid < 0)
772 			continue;
773 		node_set_online(nid);
774 	}
775 
776 	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
777 
778 	for_each_node_by_type(memory, "memory") {
779 		unsigned long start;
780 		unsigned long size;
781 		int nid;
782 		int ranges;
783 		const __be32 *memcell_buf;
784 		unsigned int len;
785 
786 		memcell_buf = of_get_property(memory,
787 			"linux,usable-memory", &len);
788 		if (!memcell_buf || len <= 0)
789 			memcell_buf = of_get_property(memory, "reg", &len);
790 		if (!memcell_buf || len <= 0)
791 			continue;
792 
793 		/* ranges in cell */
794 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
795 new_range:
796 		/* these are order-sensitive, and modify the buffer pointer */
797 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
798 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
799 
800 		/*
801 		 * Assumption: either all memory nodes or none will
802 		 * have associativity properties.  If none, then
803 		 * everything goes to default_nid.
804 		 */
805 		nid = of_node_to_nid_single(memory);
806 		if (nid < 0)
807 			nid = default_nid;
808 
809 		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
810 		node_set_online(nid);
811 
812 		if (!(size = numa_enforce_memory_limit(start, size))) {
813 			if (--ranges)
814 				goto new_range;
815 			else
816 				continue;
817 		}
818 
819 		memblock_set_node(start, size, &memblock.memory, nid);
820 
821 		if (--ranges)
822 			goto new_range;
823 	}
824 
825 	/*
826 	 * Now do the same thing for each MEMBLOCK listed in the
827 	 * ibm,dynamic-memory property in the
828 	 * ibm,dynamic-reconfiguration-memory node.
829 	 */
830 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
831 	if (memory)
832 		parse_drconf_memory(memory);
833 
834 	return 0;
835 }
836 
837 static void __init setup_nonnuma(void)
838 {
839 	unsigned long top_of_ram = memblock_end_of_DRAM();
840 	unsigned long total_ram = memblock_phys_mem_size();
841 	unsigned long start_pfn, end_pfn;
842 	unsigned int nid = 0;
843 	struct memblock_region *reg;
844 
845 	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
846 	       top_of_ram, total_ram);
847 	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
848 	       (top_of_ram - total_ram) >> 20);
849 
850 	for_each_memblock(memory, reg) {
851 		start_pfn = memblock_region_memory_base_pfn(reg);
852 		end_pfn = memblock_region_memory_end_pfn(reg);
853 
854 		fake_numa_create_new_node(end_pfn, &nid);
855 		memblock_set_node(PFN_PHYS(start_pfn),
856 				  PFN_PHYS(end_pfn - start_pfn),
857 				  &memblock.memory, nid);
858 		node_set_online(nid);
859 	}
860 }
861 
862 void __init dump_numa_cpu_topology(void)
863 {
864 	unsigned int node;
865 	unsigned int cpu, count;
866 
867 	if (min_common_depth == -1 || !numa_enabled)
868 		return;
869 
870 	for_each_online_node(node) {
871 		printk(KERN_DEBUG "Node %d CPUs:", node);
872 
873 		count = 0;
874 		/*
875 		 * If we used a CPU iterator here we would miss printing
876 		 * the holes in the cpumap.
877 		 */
878 		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
879 			if (cpumask_test_cpu(cpu,
880 					node_to_cpumask_map[node])) {
881 				if (count == 0)
882 					printk(" %u", cpu);
883 				++count;
884 			} else {
885 				if (count > 1)
886 					printk("-%u", cpu - 1);
887 				count = 0;
888 			}
889 		}
890 
891 		if (count > 1)
892 			printk("-%u", nr_cpu_ids - 1);
893 		printk("\n");
894 	}
895 }
896 
897 static void __init dump_numa_memory_topology(void)
898 {
899 	unsigned int node;
900 	unsigned int count;
901 
902 	if (min_common_depth == -1 || !numa_enabled)
903 		return;
904 
905 	for_each_online_node(node) {
906 		unsigned long i;
907 
908 		printk(KERN_DEBUG "Node %d Memory:", node);
909 
910 		count = 0;
911 
912 		for (i = 0; i < memblock_end_of_DRAM();
913 		     i += (1 << SECTION_SIZE_BITS)) {
914 			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
915 				if (count == 0)
916 					printk(" 0x%lx", i);
917 				++count;
918 			} else {
919 				if (count > 0)
920 					printk("-0x%lx", i);
921 				count = 0;
922 			}
923 		}
924 
925 		if (count > 0)
926 			printk("-0x%lx", i);
927 		printk("\n");
928 	}
929 }
930 
931 /*
932  * Allocate some memory, satisfying the memblock or bootmem allocator where
933  * required. nid is the preferred node and end is the physical address of
934  * the highest address in the node.
935  *
936  * Returns the virtual address of the memory.
937  */
938 static void __init *careful_zallocation(int nid, unsigned long size,
939 				       unsigned long align,
940 				       unsigned long end_pfn)
941 {
942 	void *ret;
943 	int new_nid;
944 	unsigned long ret_paddr;
945 
946 	ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
947 
948 	/* retry over all memory */
949 	if (!ret_paddr)
950 		ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
951 
952 	if (!ret_paddr)
953 		panic("numa.c: cannot allocate %lu bytes for node %d",
954 		      size, nid);
955 
956 	ret = __va(ret_paddr);
957 
958 	/*
959 	 * We initialize the nodes in numeric order: 0, 1, 2...
960 	 * and hand over control from the MEMBLOCK allocator to the
961 	 * bootmem allocator.  If this function is called for
962 	 * node 5, then we know that all nodes <5 are using the
963 	 * bootmem allocator instead of the MEMBLOCK allocator.
964 	 *
965 	 * So, check the nid from which this allocation came
966 	 * and double check to see if we need to use bootmem
967 	 * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
968 	 * since it would be useless.
969 	 */
970 	new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
971 	if (new_nid < nid) {
972 		ret = __alloc_bootmem_node(NODE_DATA(new_nid),
973 				size, align, 0);
974 
975 		dbg("alloc_bootmem %p %lx\n", ret, size);
976 	}
977 
978 	memset(ret, 0, size);
979 	return ret;
980 }
981 
982 static struct notifier_block ppc64_numa_nb = {
983 	.notifier_call = cpu_numa_callback,
984 	.priority = 1 /* Must run before sched domains notifier. */
985 };
986 
987 static void __init mark_reserved_regions_for_nid(int nid)
988 {
989 	struct pglist_data *node = NODE_DATA(nid);
990 	struct memblock_region *reg;
991 
992 	for_each_memblock(reserved, reg) {
993 		unsigned long physbase = reg->base;
994 		unsigned long size = reg->size;
995 		unsigned long start_pfn = physbase >> PAGE_SHIFT;
996 		unsigned long end_pfn = PFN_UP(physbase + size);
997 		struct node_active_region node_ar;
998 		unsigned long node_end_pfn = pgdat_end_pfn(node);
999 
1000 		/*
1001 		 * Check to make sure that this memblock.reserved area is
1002 		 * within the bounds of the node that we care about.
1003 		 * Checking the nid of the start and end points is not
1004 		 * sufficient because the reserved area could span the
1005 		 * entire node.
1006 		 */
1007 		if (end_pfn <= node->node_start_pfn ||
1008 		    start_pfn >= node_end_pfn)
1009 			continue;
1010 
1011 		get_node_active_region(start_pfn, &node_ar);
1012 		while (start_pfn < end_pfn &&
1013 			node_ar.start_pfn < node_ar.end_pfn) {
1014 			unsigned long reserve_size = size;
1015 			/*
1016 			 * if reserved region extends past active region
1017 			 * then trim size to active region
1018 			 */
1019 			if (end_pfn > node_ar.end_pfn)
1020 				reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1021 					- physbase;
1022 			/*
1023 			 * Only worry about *this* node, others may not
1024 			 * yet have valid NODE_DATA().
1025 			 */
1026 			if (node_ar.nid == nid) {
1027 				dbg("reserve_bootmem %lx %lx nid=%d\n",
1028 					physbase, reserve_size, node_ar.nid);
1029 				reserve_bootmem_node(NODE_DATA(node_ar.nid),
1030 						physbase, reserve_size,
1031 						BOOTMEM_DEFAULT);
1032 			}
1033 			/*
1034 			 * if reserved region is contained in the active region
1035 			 * then done.
1036 			 */
1037 			if (end_pfn <= node_ar.end_pfn)
1038 				break;
1039 
1040 			/*
1041 			 * reserved region extends past the active region
1042 			 *   get next active region that contains this
1043 			 *   reserved region
1044 			 */
1045 			start_pfn = node_ar.end_pfn;
1046 			physbase = start_pfn << PAGE_SHIFT;
1047 			size = size - reserve_size;
1048 			get_node_active_region(start_pfn, &node_ar);
1049 		}
1050 	}
1051 }
1052 
1053 
1054 void __init do_init_bootmem(void)
1055 {
1056 	int nid, cpu;
1057 
1058 	min_low_pfn = 0;
1059 	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1060 	max_pfn = max_low_pfn;
1061 
1062 	if (parse_numa_properties())
1063 		setup_nonnuma();
1064 	else
1065 		dump_numa_memory_topology();
1066 
1067 	for_each_online_node(nid) {
1068 		unsigned long start_pfn, end_pfn;
1069 		void *bootmem_vaddr;
1070 		unsigned long bootmap_pages;
1071 
1072 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1073 
1074 		/*
1075 		 * Allocate the node structure node local if possible
1076 		 *
1077 		 * Be careful moving this around, as it relies on all
1078 		 * previous nodes' bootmem to be initialized and have
1079 		 * all reserved areas marked.
1080 		 */
1081 		NODE_DATA(nid) = careful_zallocation(nid,
1082 					sizeof(struct pglist_data),
1083 					SMP_CACHE_BYTES, end_pfn);
1084 
1085   		dbg("node %d\n", nid);
1086 		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1087 
1088 		NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1089 		NODE_DATA(nid)->node_start_pfn = start_pfn;
1090 		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1091 
1092 		if (NODE_DATA(nid)->node_spanned_pages == 0)
1093   			continue;
1094 
1095   		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1096   		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1097 
1098 		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1099 		bootmem_vaddr = careful_zallocation(nid,
1100 					bootmap_pages << PAGE_SHIFT,
1101 					PAGE_SIZE, end_pfn);
1102 
1103 		dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1104 
1105 		init_bootmem_node(NODE_DATA(nid),
1106 				  __pa(bootmem_vaddr) >> PAGE_SHIFT,
1107 				  start_pfn, end_pfn);
1108 
1109 		free_bootmem_with_active_regions(nid, end_pfn);
1110 		/*
1111 		 * Be very careful about moving this around.  Future
1112 		 * calls to careful_zallocation() depend on this getting
1113 		 * done correctly.
1114 		 */
1115 		mark_reserved_regions_for_nid(nid);
1116 		sparse_memory_present_with_active_regions(nid);
1117 	}
1118 
1119 	init_bootmem_done = 1;
1120 
1121 	/*
1122 	 * Now bootmem is initialised we can create the node to cpumask
1123 	 * lookup tables and setup the cpu callback to populate them.
1124 	 */
1125 	setup_node_to_cpumask_map();
1126 
1127 	reset_numa_cpu_lookup_table();
1128 	register_cpu_notifier(&ppc64_numa_nb);
1129 	/*
1130 	 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1131 	 * even before we online them, so that we can use cpu_to_{node,mem}
1132 	 * early in boot, cf. smp_prepare_cpus().
1133 	 */
1134 	for_each_present_cpu(cpu) {
1135 		numa_setup_cpu((unsigned long)cpu);
1136 	}
1137 }
1138 
1139 static int __init early_numa(char *p)
1140 {
1141 	if (!p)
1142 		return 0;
1143 
1144 	if (strstr(p, "off"))
1145 		numa_enabled = 0;
1146 
1147 	if (strstr(p, "debug"))
1148 		numa_debug = 1;
1149 
1150 	p = strstr(p, "fake=");
1151 	if (p)
1152 		cmdline = p + strlen("fake=");
1153 
1154 	return 0;
1155 }
1156 early_param("numa", early_numa);
1157 
1158 static bool topology_updates_enabled = true;
1159 
1160 static int __init early_topology_updates(char *p)
1161 {
1162 	if (!p)
1163 		return 0;
1164 
1165 	if (!strcmp(p, "off")) {
1166 		pr_info("Disabling topology updates\n");
1167 		topology_updates_enabled = false;
1168 	}
1169 
1170 	return 0;
1171 }
1172 early_param("topology_updates", early_topology_updates);
1173 
1174 #ifdef CONFIG_MEMORY_HOTPLUG
1175 /*
1176  * Find the node associated with a hot added memory section for
1177  * memory represented in the device tree by the property
1178  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1179  */
1180 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1181 				     unsigned long scn_addr)
1182 {
1183 	const __be32 *dm;
1184 	unsigned int drconf_cell_cnt, rc;
1185 	unsigned long lmb_size;
1186 	struct assoc_arrays aa;
1187 	int nid = -1;
1188 
1189 	drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1190 	if (!drconf_cell_cnt)
1191 		return -1;
1192 
1193 	lmb_size = of_get_lmb_size(memory);
1194 	if (!lmb_size)
1195 		return -1;
1196 
1197 	rc = of_get_assoc_arrays(memory, &aa);
1198 	if (rc)
1199 		return -1;
1200 
1201 	for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1202 		struct of_drconf_cell drmem;
1203 
1204 		read_drconf_cell(&drmem, &dm);
1205 
1206 		/* skip this block if it is reserved or not assigned to
1207 		 * this partition */
1208 		if ((drmem.flags & DRCONF_MEM_RESERVED)
1209 		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1210 			continue;
1211 
1212 		if ((scn_addr < drmem.base_addr)
1213 		    || (scn_addr >= (drmem.base_addr + lmb_size)))
1214 			continue;
1215 
1216 		nid = of_drconf_to_nid_single(&drmem, &aa);
1217 		break;
1218 	}
1219 
1220 	return nid;
1221 }
1222 
1223 /*
1224  * Find the node associated with a hot added memory section for memory
1225  * represented in the device tree as a node (i.e. memory@XXXX) for
1226  * each memblock.
1227  */
1228 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1229 {
1230 	struct device_node *memory;
1231 	int nid = -1;
1232 
1233 	for_each_node_by_type(memory, "memory") {
1234 		unsigned long start, size;
1235 		int ranges;
1236 		const __be32 *memcell_buf;
1237 		unsigned int len;
1238 
1239 		memcell_buf = of_get_property(memory, "reg", &len);
1240 		if (!memcell_buf || len <= 0)
1241 			continue;
1242 
1243 		/* ranges in cell */
1244 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1245 
1246 		while (ranges--) {
1247 			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1248 			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1249 
1250 			if ((scn_addr < start) || (scn_addr >= (start + size)))
1251 				continue;
1252 
1253 			nid = of_node_to_nid_single(memory);
1254 			break;
1255 		}
1256 
1257 		if (nid >= 0)
1258 			break;
1259 	}
1260 
1261 	of_node_put(memory);
1262 
1263 	return nid;
1264 }
1265 
1266 /*
1267  * Find the node associated with a hot added memory section.  Section
1268  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1269  * sections are fully contained within a single MEMBLOCK.
1270  */
1271 int hot_add_scn_to_nid(unsigned long scn_addr)
1272 {
1273 	struct device_node *memory = NULL;
1274 	int nid, found = 0;
1275 
1276 	if (!numa_enabled || (min_common_depth < 0))
1277 		return first_online_node;
1278 
1279 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1280 	if (memory) {
1281 		nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1282 		of_node_put(memory);
1283 	} else {
1284 		nid = hot_add_node_scn_to_nid(scn_addr);
1285 	}
1286 
1287 	if (nid < 0 || !node_online(nid))
1288 		nid = first_online_node;
1289 
1290 	if (NODE_DATA(nid)->node_spanned_pages)
1291 		return nid;
1292 
1293 	for_each_online_node(nid) {
1294 		if (NODE_DATA(nid)->node_spanned_pages) {
1295 			found = 1;
1296 			break;
1297 		}
1298 	}
1299 
1300 	BUG_ON(!found);
1301 	return nid;
1302 }
1303 
1304 static u64 hot_add_drconf_memory_max(void)
1305 {
1306         struct device_node *memory = NULL;
1307         unsigned int drconf_cell_cnt = 0;
1308         u64 lmb_size = 0;
1309 	const __be32 *dm = NULL;
1310 
1311         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1312         if (memory) {
1313                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1314                 lmb_size = of_get_lmb_size(memory);
1315                 of_node_put(memory);
1316         }
1317         return lmb_size * drconf_cell_cnt;
1318 }
1319 
1320 /*
1321  * memory_hotplug_max - return max address of memory that may be added
1322  *
1323  * This is currently only used on systems that support drconfig memory
1324  * hotplug.
1325  */
1326 u64 memory_hotplug_max(void)
1327 {
1328         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1329 }
1330 #endif /* CONFIG_MEMORY_HOTPLUG */
1331 
1332 /* Virtual Processor Home Node (VPHN) support */
1333 #ifdef CONFIG_PPC_SPLPAR
1334 struct topology_update_data {
1335 	struct topology_update_data *next;
1336 	unsigned int cpu;
1337 	int old_nid;
1338 	int new_nid;
1339 };
1340 
1341 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1342 static cpumask_t cpu_associativity_changes_mask;
1343 static int vphn_enabled;
1344 static int prrn_enabled;
1345 static void reset_topology_timer(void);
1346 
1347 /*
1348  * Store the current values of the associativity change counters in the
1349  * hypervisor.
1350  */
1351 static void setup_cpu_associativity_change_counters(void)
1352 {
1353 	int cpu;
1354 
1355 	/* The VPHN feature supports a maximum of 8 reference points */
1356 	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1357 
1358 	for_each_possible_cpu(cpu) {
1359 		int i;
1360 		u8 *counts = vphn_cpu_change_counts[cpu];
1361 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1362 
1363 		for (i = 0; i < distance_ref_points_depth; i++)
1364 			counts[i] = hypervisor_counts[i];
1365 	}
1366 }
1367 
1368 /*
1369  * The hypervisor maintains a set of 8 associativity change counters in
1370  * the VPA of each cpu that correspond to the associativity levels in the
1371  * ibm,associativity-reference-points property. When an associativity
1372  * level changes, the corresponding counter is incremented.
1373  *
1374  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1375  * node associativity levels have changed.
1376  *
1377  * Returns the number of cpus with unhandled associativity changes.
1378  */
1379 static int update_cpu_associativity_changes_mask(void)
1380 {
1381 	int cpu;
1382 	cpumask_t *changes = &cpu_associativity_changes_mask;
1383 
1384 	for_each_possible_cpu(cpu) {
1385 		int i, changed = 0;
1386 		u8 *counts = vphn_cpu_change_counts[cpu];
1387 		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1388 
1389 		for (i = 0; i < distance_ref_points_depth; i++) {
1390 			if (hypervisor_counts[i] != counts[i]) {
1391 				counts[i] = hypervisor_counts[i];
1392 				changed = 1;
1393 			}
1394 		}
1395 		if (changed) {
1396 			cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1397 			cpu = cpu_last_thread_sibling(cpu);
1398 		}
1399 	}
1400 
1401 	return cpumask_weight(changes);
1402 }
1403 
1404 /*
1405  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1406  * the complete property we have to add the length in the first cell.
1407  */
1408 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1409 
1410 /*
1411  * Convert the associativity domain numbers returned from the hypervisor
1412  * to the sequence they would appear in the ibm,associativity property.
1413  */
1414 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1415 {
1416 	int i, nr_assoc_doms = 0;
1417 	const __be16 *field = (const __be16 *) packed;
1418 
1419 #define VPHN_FIELD_UNUSED	(0xffff)
1420 #define VPHN_FIELD_MSB		(0x8000)
1421 #define VPHN_FIELD_MASK		(~VPHN_FIELD_MSB)
1422 
1423 	for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1424 		if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1425 			/* All significant fields processed, and remaining
1426 			 * fields contain the reserved value of all 1's.
1427 			 * Just store them.
1428 			 */
1429 			unpacked[i] = *((__be32 *)field);
1430 			field += 2;
1431 		} else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1432 			/* Data is in the lower 15 bits of this field */
1433 			unpacked[i] = cpu_to_be32(
1434 				be16_to_cpup(field) & VPHN_FIELD_MASK);
1435 			field++;
1436 			nr_assoc_doms++;
1437 		} else {
1438 			/* Data is in the lower 15 bits of this field
1439 			 * concatenated with the next 16 bit field
1440 			 */
1441 			unpacked[i] = *((__be32 *)field);
1442 			field += 2;
1443 			nr_assoc_doms++;
1444 		}
1445 	}
1446 
1447 	/* The first cell contains the length of the property */
1448 	unpacked[0] = cpu_to_be32(nr_assoc_doms);
1449 
1450 	return nr_assoc_doms;
1451 }
1452 
1453 /*
1454  * Retrieve the new associativity information for a virtual processor's
1455  * home node.
1456  */
1457 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1458 {
1459 	long rc;
1460 	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1461 	u64 flags = 1;
1462 	int hwcpu = get_hard_smp_processor_id(cpu);
1463 	int i;
1464 
1465 	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1466 	for (i = 0; i < 6; i++)
1467 		retbuf[i] = cpu_to_be64(retbuf[i]);
1468 	vphn_unpack_associativity(retbuf, associativity);
1469 
1470 	return rc;
1471 }
1472 
1473 static long vphn_get_associativity(unsigned long cpu,
1474 					__be32 *associativity)
1475 {
1476 	long rc;
1477 
1478 	rc = hcall_vphn(cpu, associativity);
1479 
1480 	switch (rc) {
1481 	case H_FUNCTION:
1482 		printk(KERN_INFO
1483 			"VPHN is not supported. Disabling polling...\n");
1484 		stop_topology_update();
1485 		break;
1486 	case H_HARDWARE:
1487 		printk(KERN_ERR
1488 			"hcall_vphn() experienced a hardware fault "
1489 			"preventing VPHN. Disabling polling...\n");
1490 		stop_topology_update();
1491 	}
1492 
1493 	return rc;
1494 }
1495 
1496 /*
1497  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1498  * characteristics change. This function doesn't perform any locking and is
1499  * only safe to call from stop_machine().
1500  */
1501 static int update_cpu_topology(void *data)
1502 {
1503 	struct topology_update_data *update;
1504 	unsigned long cpu;
1505 
1506 	if (!data)
1507 		return -EINVAL;
1508 
1509 	cpu = smp_processor_id();
1510 
1511 	for (update = data; update; update = update->next) {
1512 		int new_nid = update->new_nid;
1513 		if (cpu != update->cpu)
1514 			continue;
1515 
1516 		unmap_cpu_from_node(cpu);
1517 		map_cpu_to_node(cpu, new_nid);
1518 		set_cpu_numa_node(cpu, new_nid);
1519 		set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1520 		vdso_getcpu_init();
1521 	}
1522 
1523 	return 0;
1524 }
1525 
1526 static int update_lookup_table(void *data)
1527 {
1528 	struct topology_update_data *update;
1529 
1530 	if (!data)
1531 		return -EINVAL;
1532 
1533 	/*
1534 	 * Upon topology update, the numa-cpu lookup table needs to be updated
1535 	 * for all threads in the core, including offline CPUs, to ensure that
1536 	 * future hotplug operations respect the cpu-to-node associativity
1537 	 * properly.
1538 	 */
1539 	for (update = data; update; update = update->next) {
1540 		int nid, base, j;
1541 
1542 		nid = update->new_nid;
1543 		base = cpu_first_thread_sibling(update->cpu);
1544 
1545 		for (j = 0; j < threads_per_core; j++) {
1546 			update_numa_cpu_lookup_table(base + j, nid);
1547 		}
1548 	}
1549 
1550 	return 0;
1551 }
1552 
1553 /*
1554  * Update the node maps and sysfs entries for each cpu whose home node
1555  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1556  */
1557 int arch_update_cpu_topology(void)
1558 {
1559 	unsigned int cpu, sibling, changed = 0;
1560 	struct topology_update_data *updates, *ud;
1561 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1562 	cpumask_t updated_cpus;
1563 	struct device *dev;
1564 	int weight, new_nid, i = 0;
1565 
1566 	if (!prrn_enabled && !vphn_enabled)
1567 		return 0;
1568 
1569 	weight = cpumask_weight(&cpu_associativity_changes_mask);
1570 	if (!weight)
1571 		return 0;
1572 
1573 	updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1574 	if (!updates)
1575 		return 0;
1576 
1577 	cpumask_clear(&updated_cpus);
1578 
1579 	for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1580 		/*
1581 		 * If siblings aren't flagged for changes, updates list
1582 		 * will be too short. Skip on this update and set for next
1583 		 * update.
1584 		 */
1585 		if (!cpumask_subset(cpu_sibling_mask(cpu),
1586 					&cpu_associativity_changes_mask)) {
1587 			pr_info("Sibling bits not set for associativity "
1588 					"change, cpu%d\n", cpu);
1589 			cpumask_or(&cpu_associativity_changes_mask,
1590 					&cpu_associativity_changes_mask,
1591 					cpu_sibling_mask(cpu));
1592 			cpu = cpu_last_thread_sibling(cpu);
1593 			continue;
1594 		}
1595 
1596 		/* Use associativity from first thread for all siblings */
1597 		vphn_get_associativity(cpu, associativity);
1598 		new_nid = associativity_to_nid(associativity);
1599 		if (new_nid < 0 || !node_online(new_nid))
1600 			new_nid = first_online_node;
1601 
1602 		if (new_nid == numa_cpu_lookup_table[cpu]) {
1603 			cpumask_andnot(&cpu_associativity_changes_mask,
1604 					&cpu_associativity_changes_mask,
1605 					cpu_sibling_mask(cpu));
1606 			cpu = cpu_last_thread_sibling(cpu);
1607 			continue;
1608 		}
1609 
1610 		for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1611 			ud = &updates[i++];
1612 			ud->cpu = sibling;
1613 			ud->new_nid = new_nid;
1614 			ud->old_nid = numa_cpu_lookup_table[sibling];
1615 			cpumask_set_cpu(sibling, &updated_cpus);
1616 			if (i < weight)
1617 				ud->next = &updates[i];
1618 		}
1619 		cpu = cpu_last_thread_sibling(cpu);
1620 	}
1621 
1622 	pr_debug("Topology update for the following CPUs:\n");
1623 	if (cpumask_weight(&updated_cpus)) {
1624 		for (ud = &updates[0]; ud; ud = ud->next) {
1625 			pr_debug("cpu %d moving from node %d "
1626 					  "to %d\n", ud->cpu,
1627 					  ud->old_nid, ud->new_nid);
1628 		}
1629 	}
1630 
1631 	/*
1632 	 * In cases where we have nothing to update (because the updates list
1633 	 * is too short or because the new topology is same as the old one),
1634 	 * skip invoking update_cpu_topology() via stop-machine(). This is
1635 	 * necessary (and not just a fast-path optimization) since stop-machine
1636 	 * can end up electing a random CPU to run update_cpu_topology(), and
1637 	 * thus trick us into setting up incorrect cpu-node mappings (since
1638 	 * 'updates' is kzalloc()'ed).
1639 	 *
1640 	 * And for the similar reason, we will skip all the following updating.
1641 	 */
1642 	if (!cpumask_weight(&updated_cpus))
1643 		goto out;
1644 
1645 	stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1646 
1647 	/*
1648 	 * Update the numa-cpu lookup table with the new mappings, even for
1649 	 * offline CPUs. It is best to perform this update from the stop-
1650 	 * machine context.
1651 	 */
1652 	stop_machine(update_lookup_table, &updates[0],
1653 					cpumask_of(raw_smp_processor_id()));
1654 
1655 	for (ud = &updates[0]; ud; ud = ud->next) {
1656 		unregister_cpu_under_node(ud->cpu, ud->old_nid);
1657 		register_cpu_under_node(ud->cpu, ud->new_nid);
1658 
1659 		dev = get_cpu_device(ud->cpu);
1660 		if (dev)
1661 			kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1662 		cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1663 		changed = 1;
1664 	}
1665 
1666 out:
1667 	kfree(updates);
1668 	return changed;
1669 }
1670 
1671 static void topology_work_fn(struct work_struct *work)
1672 {
1673 	rebuild_sched_domains();
1674 }
1675 static DECLARE_WORK(topology_work, topology_work_fn);
1676 
1677 static void topology_schedule_update(void)
1678 {
1679 	schedule_work(&topology_work);
1680 }
1681 
1682 static void topology_timer_fn(unsigned long ignored)
1683 {
1684 	if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1685 		topology_schedule_update();
1686 	else if (vphn_enabled) {
1687 		if (update_cpu_associativity_changes_mask() > 0)
1688 			topology_schedule_update();
1689 		reset_topology_timer();
1690 	}
1691 }
1692 static struct timer_list topology_timer =
1693 	TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1694 
1695 static void reset_topology_timer(void)
1696 {
1697 	topology_timer.data = 0;
1698 	topology_timer.expires = jiffies + 60 * HZ;
1699 	mod_timer(&topology_timer, topology_timer.expires);
1700 }
1701 
1702 #ifdef CONFIG_SMP
1703 
1704 static void stage_topology_update(int core_id)
1705 {
1706 	cpumask_or(&cpu_associativity_changes_mask,
1707 		&cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1708 	reset_topology_timer();
1709 }
1710 
1711 static int dt_update_callback(struct notifier_block *nb,
1712 				unsigned long action, void *data)
1713 {
1714 	struct of_prop_reconfig *update;
1715 	int rc = NOTIFY_DONE;
1716 
1717 	switch (action) {
1718 	case OF_RECONFIG_UPDATE_PROPERTY:
1719 		update = (struct of_prop_reconfig *)data;
1720 		if (!of_prop_cmp(update->dn->type, "cpu") &&
1721 		    !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1722 			u32 core_id;
1723 			of_property_read_u32(update->dn, "reg", &core_id);
1724 			stage_topology_update(core_id);
1725 			rc = NOTIFY_OK;
1726 		}
1727 		break;
1728 	}
1729 
1730 	return rc;
1731 }
1732 
1733 static struct notifier_block dt_update_nb = {
1734 	.notifier_call = dt_update_callback,
1735 };
1736 
1737 #endif
1738 
1739 /*
1740  * Start polling for associativity changes.
1741  */
1742 int start_topology_update(void)
1743 {
1744 	int rc = 0;
1745 
1746 	if (firmware_has_feature(FW_FEATURE_PRRN)) {
1747 		if (!prrn_enabled) {
1748 			prrn_enabled = 1;
1749 			vphn_enabled = 0;
1750 #ifdef CONFIG_SMP
1751 			rc = of_reconfig_notifier_register(&dt_update_nb);
1752 #endif
1753 		}
1754 	} else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1755 		   lppaca_shared_proc(get_lppaca())) {
1756 		if (!vphn_enabled) {
1757 			prrn_enabled = 0;
1758 			vphn_enabled = 1;
1759 			setup_cpu_associativity_change_counters();
1760 			init_timer_deferrable(&topology_timer);
1761 			reset_topology_timer();
1762 		}
1763 	}
1764 
1765 	return rc;
1766 }
1767 
1768 /*
1769  * Disable polling for VPHN associativity changes.
1770  */
1771 int stop_topology_update(void)
1772 {
1773 	int rc = 0;
1774 
1775 	if (prrn_enabled) {
1776 		prrn_enabled = 0;
1777 #ifdef CONFIG_SMP
1778 		rc = of_reconfig_notifier_unregister(&dt_update_nb);
1779 #endif
1780 	} else if (vphn_enabled) {
1781 		vphn_enabled = 0;
1782 		rc = del_timer_sync(&topology_timer);
1783 	}
1784 
1785 	return rc;
1786 }
1787 
1788 int prrn_is_enabled(void)
1789 {
1790 	return prrn_enabled;
1791 }
1792 
1793 static int topology_read(struct seq_file *file, void *v)
1794 {
1795 	if (vphn_enabled || prrn_enabled)
1796 		seq_puts(file, "on\n");
1797 	else
1798 		seq_puts(file, "off\n");
1799 
1800 	return 0;
1801 }
1802 
1803 static int topology_open(struct inode *inode, struct file *file)
1804 {
1805 	return single_open(file, topology_read, NULL);
1806 }
1807 
1808 static ssize_t topology_write(struct file *file, const char __user *buf,
1809 			      size_t count, loff_t *off)
1810 {
1811 	char kbuf[4]; /* "on" or "off" plus null. */
1812 	int read_len;
1813 
1814 	read_len = count < 3 ? count : 3;
1815 	if (copy_from_user(kbuf, buf, read_len))
1816 		return -EINVAL;
1817 
1818 	kbuf[read_len] = '\0';
1819 
1820 	if (!strncmp(kbuf, "on", 2))
1821 		start_topology_update();
1822 	else if (!strncmp(kbuf, "off", 3))
1823 		stop_topology_update();
1824 	else
1825 		return -EINVAL;
1826 
1827 	return count;
1828 }
1829 
1830 static const struct file_operations topology_ops = {
1831 	.read = seq_read,
1832 	.write = topology_write,
1833 	.open = topology_open,
1834 	.release = single_release
1835 };
1836 
1837 static int topology_update_init(void)
1838 {
1839 	/* Do not poll for changes if disabled at boot */
1840 	if (topology_updates_enabled)
1841 		start_topology_update();
1842 
1843 	if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1844 		return -ENOMEM;
1845 
1846 	return 0;
1847 }
1848 device_initcall(topology_update_init);
1849 #endif /* CONFIG_PPC_SPLPAR */
1850