4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
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.
11 #define pr_fmt(fmt) "numa: " fmt
13 #include <linux/threads.h>
14 #include <linux/bootmem.h>
15 #include <linux/init.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>
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>
36 #include <asm/cputhreads.h>
37 #include <asm/topology.h>
38 #include <asm/firmware.h>
40 #include <asm/hvcall.h>
41 #include <asm/setup.h>
44 static int numa_enabled = 1;
46 static char *cmdline __initdata;
48 static int numa_debug;
49 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
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];
55 EXPORT_SYMBOL(numa_cpu_lookup_table);
56 EXPORT_SYMBOL(node_to_cpumask_map);
57 EXPORT_SYMBOL(node_data);
59 static int min_common_depth;
60 static int n_mem_addr_cells, n_mem_size_cells;
61 static int form1_affinity;
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];
69 * Allocate node_to_cpumask_map based on number of available nodes
70 * Requires node_possible_map to be valid.
72 * Note: cpumask_of_node() is not valid until after this is done.
74 static void __init setup_node_to_cpumask_map(void)
78 /* setup nr_node_ids if not done yet */
79 if (nr_node_ids == MAX_NUMNODES)
82 /* allocate the map */
83 for (node = 0; node < nr_node_ids; node++)
84 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
86 /* cpumask_of_node() will now work */
87 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
90 static int __init fake_numa_create_new_node(unsigned long end_pfn,
93 unsigned long long mem;
95 static unsigned int fake_nid;
96 static unsigned long long curr_boundary;
99 * Modify node id, iff we started creating NUMA nodes
100 * We want to continue from where we left of the last time
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).
112 mem = memparse(p, &p);
116 if (mem < curr_boundary)
121 if ((end_pfn << PAGE_SHIFT) > mem) {
123 * Skip commas and spaces
125 while (*p == ',' || *p == ' ' || *p == '\t')
131 dbg("created new fake_node with id %d\n", fake_nid);
137 static void reset_numa_cpu_lookup_table(void)
141 for_each_possible_cpu(cpu)
142 numa_cpu_lookup_table[cpu] = -1;
145 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
147 numa_cpu_lookup_table[cpu] = node;
150 static void map_cpu_to_node(int cpu, int node)
152 update_numa_cpu_lookup_table(cpu, node);
154 dbg("adding cpu %d to node %d\n", cpu, node);
156 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
157 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
160 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
161 static void unmap_cpu_from_node(unsigned long cpu)
163 int node = numa_cpu_lookup_table[cpu];
165 dbg("removing cpu %lu from node %d\n", cpu, node);
167 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
168 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
170 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
174 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
176 /* must hold reference to node during call */
177 static const __be32 *of_get_associativity(struct device_node *dev)
179 return of_get_property(dev, "ibm,associativity", NULL);
183 * Returns the property linux,drconf-usable-memory if
184 * it exists (the property exists only in kexec/kdump kernels,
185 * added by kexec-tools)
187 static const __be32 *of_get_usable_memory(struct device_node *memory)
191 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
192 if (!prop || len < sizeof(unsigned int))
197 int __node_distance(int a, int b)
200 int distance = LOCAL_DISTANCE;
203 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
205 for (i = 0; i < distance_ref_points_depth; i++) {
206 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
209 /* Double the distance for each NUMA level */
215 EXPORT_SYMBOL(__node_distance);
217 static void initialize_distance_lookup_table(int nid,
218 const __be32 *associativity)
225 for (i = 0; i < distance_ref_points_depth; i++) {
228 entry = &associativity[be32_to_cpu(distance_ref_points[i])];
229 distance_lookup_table[nid][i] = of_read_number(entry, 1);
233 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
236 static int associativity_to_nid(const __be32 *associativity)
240 if (min_common_depth == -1)
243 if (of_read_number(associativity, 1) >= min_common_depth)
244 nid = of_read_number(&associativity[min_common_depth], 1);
246 /* POWER4 LPAR uses 0xffff as invalid node */
247 if (nid == 0xffff || nid >= MAX_NUMNODES)
251 of_read_number(associativity, 1) >= distance_ref_points_depth)
252 initialize_distance_lookup_table(nid, associativity);
258 /* Returns the nid associated with the given device tree node,
259 * or -1 if not found.
261 static int of_node_to_nid_single(struct device_node *device)
266 tmp = of_get_associativity(device);
268 nid = associativity_to_nid(tmp);
272 /* Walk the device tree upwards, looking for an associativity id */
273 int of_node_to_nid(struct device_node *device)
275 struct device_node *tmp;
280 nid = of_node_to_nid_single(device);
285 device = of_get_parent(tmp);
292 EXPORT_SYMBOL_GPL(of_node_to_nid);
294 static int __init find_min_common_depth(void)
297 struct device_node *root;
299 if (firmware_has_feature(FW_FEATURE_OPAL))
300 root = of_find_node_by_path("/ibm,opal");
302 root = of_find_node_by_path("/rtas");
304 root = of_find_node_by_path("/");
307 * This property is a set of 32-bit integers, each representing
308 * an index into the ibm,associativity nodes.
310 * With form 0 affinity the first integer is for an SMP configuration
311 * (should be all 0's) and the second is for a normal NUMA
312 * configuration. We have only one level of NUMA.
314 * With form 1 affinity the first integer is the most significant
315 * NUMA boundary and the following are progressively less significant
316 * boundaries. There can be more than one level of NUMA.
318 distance_ref_points = of_get_property(root,
319 "ibm,associativity-reference-points",
320 &distance_ref_points_depth);
322 if (!distance_ref_points) {
323 dbg("NUMA: ibm,associativity-reference-points not found.\n");
327 distance_ref_points_depth /= sizeof(int);
329 if (firmware_has_feature(FW_FEATURE_OPAL) ||
330 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
331 dbg("Using form 1 affinity\n");
335 if (form1_affinity) {
336 depth = of_read_number(distance_ref_points, 1);
338 if (distance_ref_points_depth < 2) {
339 printk(KERN_WARNING "NUMA: "
340 "short ibm,associativity-reference-points\n");
344 depth = of_read_number(&distance_ref_points[1], 1);
348 * Warn and cap if the hardware supports more than
349 * MAX_DISTANCE_REF_POINTS domains.
351 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
352 printk(KERN_WARNING "NUMA: distance array capped at "
353 "%d entries\n", MAX_DISTANCE_REF_POINTS);
354 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
365 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
367 struct device_node *memory = NULL;
369 memory = of_find_node_by_type(memory, "memory");
371 panic("numa.c: No memory nodes found!");
373 *n_addr_cells = of_n_addr_cells(memory);
374 *n_size_cells = of_n_size_cells(memory);
378 static unsigned long read_n_cells(int n, const __be32 **buf)
380 unsigned long result = 0;
383 result = (result << 32) | of_read_number(*buf, 1);
390 * Read the next memblock list entry from the ibm,dynamic-memory property
391 * and return the information in the provided of_drconf_cell structure.
393 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
397 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
400 drmem->drc_index = of_read_number(cp, 1);
401 drmem->reserved = of_read_number(&cp[1], 1);
402 drmem->aa_index = of_read_number(&cp[2], 1);
403 drmem->flags = of_read_number(&cp[3], 1);
409 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
411 * The layout of the ibm,dynamic-memory property is a number N of memblock
412 * list entries followed by N memblock list entries. Each memblock list entry
413 * contains information as laid out in the of_drconf_cell struct above.
415 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
420 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
421 if (!prop || len < sizeof(unsigned int))
424 entries = of_read_number(prop++, 1);
426 /* Now that we know the number of entries, revalidate the size
427 * of the property read in to ensure we have everything
429 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
437 * Retrieve and validate the ibm,lmb-size property for drconf memory
438 * from the device tree.
440 static u64 of_get_lmb_size(struct device_node *memory)
445 prop = of_get_property(memory, "ibm,lmb-size", &len);
446 if (!prop || len < sizeof(unsigned int))
449 return read_n_cells(n_mem_size_cells, &prop);
452 struct assoc_arrays {
455 const __be32 *arrays;
459 * Retrieve and validate the list of associativity arrays for drconf
460 * memory from the ibm,associativity-lookup-arrays property of the
463 * The layout of the ibm,associativity-lookup-arrays property is a number N
464 * indicating the number of associativity arrays, followed by a number M
465 * indicating the size of each associativity array, followed by a list
466 * of N associativity arrays.
468 static int of_get_assoc_arrays(struct device_node *memory,
469 struct assoc_arrays *aa)
474 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
475 if (!prop || len < 2 * sizeof(unsigned int))
478 aa->n_arrays = of_read_number(prop++, 1);
479 aa->array_sz = of_read_number(prop++, 1);
481 /* Now that we know the number of arrays and size of each array,
482 * revalidate the size of the property read in.
484 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
492 * This is like of_node_to_nid_single() for memory represented in the
493 * ibm,dynamic-reconfiguration-memory node.
495 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
496 struct assoc_arrays *aa)
499 int nid = default_nid;
502 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
503 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
504 drmem->aa_index < aa->n_arrays) {
505 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
506 nid = of_read_number(&aa->arrays[index], 1);
508 if (nid == 0xffff || nid >= MAX_NUMNODES)
516 * Figure out to which domain a cpu belongs and stick it there.
517 * Return the id of the domain used.
519 static int numa_setup_cpu(unsigned long lcpu)
522 struct device_node *cpu;
525 * If a valid cpu-to-node mapping is already available, use it
526 * directly instead of querying the firmware, since it represents
527 * the most recent mapping notified to us by the platform (eg: VPHN).
529 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
530 map_cpu_to_node(lcpu, nid);
534 cpu = of_get_cpu_node(lcpu, NULL);
538 if (cpu_present(lcpu))
544 nid = of_node_to_nid_single(cpu);
547 if (nid < 0 || !node_online(nid))
548 nid = first_online_node;
550 map_cpu_to_node(lcpu, nid);
556 static void verify_cpu_node_mapping(int cpu, int node)
558 int base, sibling, i;
560 /* Verify that all the threads in the core belong to the same node */
561 base = cpu_first_thread_sibling(cpu);
563 for (i = 0; i < threads_per_core; i++) {
566 if (sibling == cpu || cpu_is_offline(sibling))
569 if (cpu_to_node(sibling) != node) {
570 WARN(1, "CPU thread siblings %d and %d don't belong"
571 " to the same node!\n", cpu, sibling);
577 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
580 unsigned long lcpu = (unsigned long)hcpu;
581 int ret = NOTIFY_DONE, nid;
585 case CPU_UP_PREPARE_FROZEN:
586 nid = numa_setup_cpu(lcpu);
587 verify_cpu_node_mapping((int)lcpu, nid);
590 #ifdef CONFIG_HOTPLUG_CPU
592 case CPU_DEAD_FROZEN:
593 case CPU_UP_CANCELED:
594 case CPU_UP_CANCELED_FROZEN:
595 unmap_cpu_from_node(lcpu);
604 * Check and possibly modify a memory region to enforce the memory limit.
606 * Returns the size the region should have to enforce the memory limit.
607 * This will either be the original value of size, a truncated value,
608 * or zero. If the returned value of size is 0 the region should be
609 * discarded as it lies wholly above the memory limit.
611 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
615 * We use memblock_end_of_DRAM() in here instead of memory_limit because
616 * we've already adjusted it for the limit and it takes care of
617 * having memory holes below the limit. Also, in the case of
618 * iommu_is_off, memory_limit is not set but is implicitly enforced.
621 if (start + size <= memblock_end_of_DRAM())
624 if (start >= memblock_end_of_DRAM())
627 return memblock_end_of_DRAM() - start;
631 * Reads the counter for a given entry in
632 * linux,drconf-usable-memory property
634 static inline int __init read_usm_ranges(const __be32 **usm)
637 * For each lmb in ibm,dynamic-memory a corresponding
638 * entry in linux,drconf-usable-memory property contains
639 * a counter followed by that many (base, size) duple.
640 * read the counter from linux,drconf-usable-memory
642 return read_n_cells(n_mem_size_cells, usm);
646 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
647 * node. This assumes n_mem_{addr,size}_cells have been set.
649 static void __init parse_drconf_memory(struct device_node *memory)
651 const __be32 *uninitialized_var(dm), *usm;
652 unsigned int n, rc, ranges, is_kexec_kdump = 0;
653 unsigned long lmb_size, base, size, sz;
655 struct assoc_arrays aa = { .arrays = NULL };
657 n = of_get_drconf_memory(memory, &dm);
661 lmb_size = of_get_lmb_size(memory);
665 rc = of_get_assoc_arrays(memory, &aa);
669 /* check if this is a kexec/kdump kernel */
670 usm = of_get_usable_memory(memory);
674 for (; n != 0; --n) {
675 struct of_drconf_cell drmem;
677 read_drconf_cell(&drmem, &dm);
679 /* skip this block if the reserved bit is set in flags (0x80)
680 or if the block is not assigned to this partition (0x8) */
681 if ((drmem.flags & DRCONF_MEM_RESERVED)
682 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
685 base = drmem.base_addr;
689 if (is_kexec_kdump) {
690 ranges = read_usm_ranges(&usm);
691 if (!ranges) /* there are no (base, size) duple */
695 if (is_kexec_kdump) {
696 base = read_n_cells(n_mem_addr_cells, &usm);
697 size = read_n_cells(n_mem_size_cells, &usm);
699 nid = of_drconf_to_nid_single(&drmem, &aa);
700 fake_numa_create_new_node(
701 ((base + size) >> PAGE_SHIFT),
703 node_set_online(nid);
704 sz = numa_enforce_memory_limit(base, size);
706 memblock_set_node(base, sz,
707 &memblock.memory, nid);
712 static int __init parse_numa_properties(void)
714 struct device_node *memory;
718 if (numa_enabled == 0) {
719 printk(KERN_WARNING "NUMA disabled by user\n");
723 min_common_depth = find_min_common_depth();
725 if (min_common_depth < 0)
726 return min_common_depth;
728 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
731 * Even though we connect cpus to numa domains later in SMP
732 * init, we need to know the node ids now. This is because
733 * each node to be onlined must have NODE_DATA etc backing it.
735 for_each_present_cpu(i) {
736 struct device_node *cpu;
739 cpu = of_get_cpu_node(i, NULL);
741 nid = of_node_to_nid_single(cpu);
745 * Don't fall back to default_nid yet -- we will plug
746 * cpus into nodes once the memory scan has discovered
751 node_set_online(nid);
754 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
756 for_each_node_by_type(memory, "memory") {
761 const __be32 *memcell_buf;
764 memcell_buf = of_get_property(memory,
765 "linux,usable-memory", &len);
766 if (!memcell_buf || len <= 0)
767 memcell_buf = of_get_property(memory, "reg", &len);
768 if (!memcell_buf || len <= 0)
772 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
774 /* these are order-sensitive, and modify the buffer pointer */
775 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
776 size = read_n_cells(n_mem_size_cells, &memcell_buf);
779 * Assumption: either all memory nodes or none will
780 * have associativity properties. If none, then
781 * everything goes to default_nid.
783 nid = of_node_to_nid_single(memory);
787 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
788 node_set_online(nid);
790 if (!(size = numa_enforce_memory_limit(start, size))) {
797 memblock_set_node(start, size, &memblock.memory, nid);
804 * Now do the same thing for each MEMBLOCK listed in the
805 * ibm,dynamic-memory property in the
806 * ibm,dynamic-reconfiguration-memory node.
808 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
810 parse_drconf_memory(memory);
815 static void __init setup_nonnuma(void)
817 unsigned long top_of_ram = memblock_end_of_DRAM();
818 unsigned long total_ram = memblock_phys_mem_size();
819 unsigned long start_pfn, end_pfn;
820 unsigned int nid = 0;
821 struct memblock_region *reg;
823 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
824 top_of_ram, total_ram);
825 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
826 (top_of_ram - total_ram) >> 20);
828 for_each_memblock(memory, reg) {
829 start_pfn = memblock_region_memory_base_pfn(reg);
830 end_pfn = memblock_region_memory_end_pfn(reg);
832 fake_numa_create_new_node(end_pfn, &nid);
833 memblock_set_node(PFN_PHYS(start_pfn),
834 PFN_PHYS(end_pfn - start_pfn),
835 &memblock.memory, nid);
836 node_set_online(nid);
840 void __init dump_numa_cpu_topology(void)
843 unsigned int cpu, count;
845 if (min_common_depth == -1 || !numa_enabled)
848 for_each_online_node(node) {
849 printk(KERN_DEBUG "Node %d CPUs:", node);
853 * If we used a CPU iterator here we would miss printing
854 * the holes in the cpumap.
856 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
857 if (cpumask_test_cpu(cpu,
858 node_to_cpumask_map[node])) {
864 printk("-%u", cpu - 1);
870 printk("-%u", nr_cpu_ids - 1);
875 static void __init dump_numa_memory_topology(void)
880 if (min_common_depth == -1 || !numa_enabled)
883 for_each_online_node(node) {
886 printk(KERN_DEBUG "Node %d Memory:", node);
890 for (i = 0; i < memblock_end_of_DRAM();
891 i += (1 << SECTION_SIZE_BITS)) {
892 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
909 static struct notifier_block ppc64_numa_nb = {
910 .notifier_call = cpu_numa_callback,
911 .priority = 1 /* Must run before sched domains notifier. */
914 /* Initialize NODE_DATA for a node on the local memory */
915 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
917 u64 spanned_pages = end_pfn - start_pfn;
918 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
924 pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
925 nid, start_pfn << PAGE_SHIFT,
926 (end_pfn << PAGE_SHIFT) - 1);
928 pr_info("Initmem setup node %d\n", nid);
930 nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
933 /* report and initialize */
934 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
935 nd_pa, nd_pa + nd_size - 1);
936 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
938 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
941 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
942 NODE_DATA(nid)->node_id = nid;
943 NODE_DATA(nid)->node_start_pfn = start_pfn;
944 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
947 void __init initmem_init(void)
951 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
952 max_pfn = max_low_pfn;
954 if (parse_numa_properties())
957 dump_numa_memory_topology();
962 * Reduce the possible NUMA nodes to the online NUMA nodes,
963 * since we do not support node hotplug. This ensures that we
964 * lower the maximum NUMA node ID to what is actually present.
966 nodes_and(node_possible_map, node_possible_map, node_online_map);
968 for_each_online_node(nid) {
969 unsigned long start_pfn, end_pfn;
971 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
972 setup_node_data(nid, start_pfn, end_pfn);
973 sparse_memory_present_with_active_regions(nid);
978 setup_node_to_cpumask_map();
980 reset_numa_cpu_lookup_table();
981 register_cpu_notifier(&ppc64_numa_nb);
983 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
984 * even before we online them, so that we can use cpu_to_{node,mem}
985 * early in boot, cf. smp_prepare_cpus().
987 for_each_present_cpu(cpu) {
988 numa_setup_cpu((unsigned long)cpu);
992 static int __init early_numa(char *p)
997 if (strstr(p, "off"))
1000 if (strstr(p, "debug"))
1003 p = strstr(p, "fake=");
1005 cmdline = p + strlen("fake=");
1009 early_param("numa", early_numa);
1011 static bool topology_updates_enabled = true;
1013 static int __init early_topology_updates(char *p)
1018 if (!strcmp(p, "off")) {
1019 pr_info("Disabling topology updates\n");
1020 topology_updates_enabled = false;
1025 early_param("topology_updates", early_topology_updates);
1027 #ifdef CONFIG_MEMORY_HOTPLUG
1029 * Find the node associated with a hot added memory section for
1030 * memory represented in the device tree by the property
1031 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1033 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1034 unsigned long scn_addr)
1037 unsigned int drconf_cell_cnt, rc;
1038 unsigned long lmb_size;
1039 struct assoc_arrays aa;
1042 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1043 if (!drconf_cell_cnt)
1046 lmb_size = of_get_lmb_size(memory);
1050 rc = of_get_assoc_arrays(memory, &aa);
1054 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1055 struct of_drconf_cell drmem;
1057 read_drconf_cell(&drmem, &dm);
1059 /* skip this block if it is reserved or not assigned to
1061 if ((drmem.flags & DRCONF_MEM_RESERVED)
1062 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1065 if ((scn_addr < drmem.base_addr)
1066 || (scn_addr >= (drmem.base_addr + lmb_size)))
1069 nid = of_drconf_to_nid_single(&drmem, &aa);
1077 * Find the node associated with a hot added memory section for memory
1078 * represented in the device tree as a node (i.e. memory@XXXX) for
1081 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1083 struct device_node *memory;
1086 for_each_node_by_type(memory, "memory") {
1087 unsigned long start, size;
1089 const __be32 *memcell_buf;
1092 memcell_buf = of_get_property(memory, "reg", &len);
1093 if (!memcell_buf || len <= 0)
1096 /* ranges in cell */
1097 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1100 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1101 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1103 if ((scn_addr < start) || (scn_addr >= (start + size)))
1106 nid = of_node_to_nid_single(memory);
1114 of_node_put(memory);
1120 * Find the node associated with a hot added memory section. Section
1121 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1122 * sections are fully contained within a single MEMBLOCK.
1124 int hot_add_scn_to_nid(unsigned long scn_addr)
1126 struct device_node *memory = NULL;
1129 if (!numa_enabled || (min_common_depth < 0))
1130 return first_online_node;
1132 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1134 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1135 of_node_put(memory);
1137 nid = hot_add_node_scn_to_nid(scn_addr);
1140 if (nid < 0 || !node_online(nid))
1141 nid = first_online_node;
1143 if (NODE_DATA(nid)->node_spanned_pages)
1146 for_each_online_node(nid) {
1147 if (NODE_DATA(nid)->node_spanned_pages) {
1157 static u64 hot_add_drconf_memory_max(void)
1159 struct device_node *memory = NULL;
1160 unsigned int drconf_cell_cnt = 0;
1162 const __be32 *dm = NULL;
1164 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1166 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1167 lmb_size = of_get_lmb_size(memory);
1168 of_node_put(memory);
1170 return lmb_size * drconf_cell_cnt;
1174 * memory_hotplug_max - return max address of memory that may be added
1176 * This is currently only used on systems that support drconfig memory
1179 u64 memory_hotplug_max(void)
1181 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1183 #endif /* CONFIG_MEMORY_HOTPLUG */
1185 /* Virtual Processor Home Node (VPHN) support */
1186 #ifdef CONFIG_PPC_SPLPAR
1190 struct topology_update_data {
1191 struct topology_update_data *next;
1197 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1198 static cpumask_t cpu_associativity_changes_mask;
1199 static int vphn_enabled;
1200 static int prrn_enabled;
1201 static void reset_topology_timer(void);
1204 * Store the current values of the associativity change counters in the
1207 static void setup_cpu_associativity_change_counters(void)
1211 /* The VPHN feature supports a maximum of 8 reference points */
1212 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1214 for_each_possible_cpu(cpu) {
1216 u8 *counts = vphn_cpu_change_counts[cpu];
1217 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1219 for (i = 0; i < distance_ref_points_depth; i++)
1220 counts[i] = hypervisor_counts[i];
1225 * The hypervisor maintains a set of 8 associativity change counters in
1226 * the VPA of each cpu that correspond to the associativity levels in the
1227 * ibm,associativity-reference-points property. When an associativity
1228 * level changes, the corresponding counter is incremented.
1230 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1231 * node associativity levels have changed.
1233 * Returns the number of cpus with unhandled associativity changes.
1235 static int update_cpu_associativity_changes_mask(void)
1238 cpumask_t *changes = &cpu_associativity_changes_mask;
1240 for_each_possible_cpu(cpu) {
1242 u8 *counts = vphn_cpu_change_counts[cpu];
1243 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1245 for (i = 0; i < distance_ref_points_depth; i++) {
1246 if (hypervisor_counts[i] != counts[i]) {
1247 counts[i] = hypervisor_counts[i];
1252 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1253 cpu = cpu_last_thread_sibling(cpu);
1257 return cpumask_weight(changes);
1261 * Retrieve the new associativity information for a virtual processor's
1264 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1267 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1269 int hwcpu = get_hard_smp_processor_id(cpu);
1271 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1272 vphn_unpack_associativity(retbuf, associativity);
1277 static long vphn_get_associativity(unsigned long cpu,
1278 __be32 *associativity)
1282 rc = hcall_vphn(cpu, associativity);
1287 "VPHN is not supported. Disabling polling...\n");
1288 stop_topology_update();
1292 "hcall_vphn() experienced a hardware fault "
1293 "preventing VPHN. Disabling polling...\n");
1294 stop_topology_update();
1301 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1302 * characteristics change. This function doesn't perform any locking and is
1303 * only safe to call from stop_machine().
1305 static int update_cpu_topology(void *data)
1307 struct topology_update_data *update;
1313 cpu = smp_processor_id();
1315 for (update = data; update; update = update->next) {
1316 int new_nid = update->new_nid;
1317 if (cpu != update->cpu)
1320 unmap_cpu_from_node(cpu);
1321 map_cpu_to_node(cpu, new_nid);
1322 set_cpu_numa_node(cpu, new_nid);
1323 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1330 static int update_lookup_table(void *data)
1332 struct topology_update_data *update;
1338 * Upon topology update, the numa-cpu lookup table needs to be updated
1339 * for all threads in the core, including offline CPUs, to ensure that
1340 * future hotplug operations respect the cpu-to-node associativity
1343 for (update = data; update; update = update->next) {
1346 nid = update->new_nid;
1347 base = cpu_first_thread_sibling(update->cpu);
1349 for (j = 0; j < threads_per_core; j++) {
1350 update_numa_cpu_lookup_table(base + j, nid);
1358 * Update the node maps and sysfs entries for each cpu whose home node
1359 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1361 int arch_update_cpu_topology(void)
1363 unsigned int cpu, sibling, changed = 0;
1364 struct topology_update_data *updates, *ud;
1365 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1366 cpumask_t updated_cpus;
1368 int weight, new_nid, i = 0;
1370 if (!prrn_enabled && !vphn_enabled)
1373 weight = cpumask_weight(&cpu_associativity_changes_mask);
1377 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1381 cpumask_clear(&updated_cpus);
1383 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1385 * If siblings aren't flagged for changes, updates list
1386 * will be too short. Skip on this update and set for next
1389 if (!cpumask_subset(cpu_sibling_mask(cpu),
1390 &cpu_associativity_changes_mask)) {
1391 pr_info("Sibling bits not set for associativity "
1392 "change, cpu%d\n", cpu);
1393 cpumask_or(&cpu_associativity_changes_mask,
1394 &cpu_associativity_changes_mask,
1395 cpu_sibling_mask(cpu));
1396 cpu = cpu_last_thread_sibling(cpu);
1400 /* Use associativity from first thread for all siblings */
1401 vphn_get_associativity(cpu, associativity);
1402 new_nid = associativity_to_nid(associativity);
1403 if (new_nid < 0 || !node_online(new_nid))
1404 new_nid = first_online_node;
1406 if (new_nid == numa_cpu_lookup_table[cpu]) {
1407 cpumask_andnot(&cpu_associativity_changes_mask,
1408 &cpu_associativity_changes_mask,
1409 cpu_sibling_mask(cpu));
1410 cpu = cpu_last_thread_sibling(cpu);
1414 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1417 ud->new_nid = new_nid;
1418 ud->old_nid = numa_cpu_lookup_table[sibling];
1419 cpumask_set_cpu(sibling, &updated_cpus);
1421 ud->next = &updates[i];
1423 cpu = cpu_last_thread_sibling(cpu);
1426 pr_debug("Topology update for the following CPUs:\n");
1427 if (cpumask_weight(&updated_cpus)) {
1428 for (ud = &updates[0]; ud; ud = ud->next) {
1429 pr_debug("cpu %d moving from node %d "
1431 ud->old_nid, ud->new_nid);
1436 * In cases where we have nothing to update (because the updates list
1437 * is too short or because the new topology is same as the old one),
1438 * skip invoking update_cpu_topology() via stop-machine(). This is
1439 * necessary (and not just a fast-path optimization) since stop-machine
1440 * can end up electing a random CPU to run update_cpu_topology(), and
1441 * thus trick us into setting up incorrect cpu-node mappings (since
1442 * 'updates' is kzalloc()'ed).
1444 * And for the similar reason, we will skip all the following updating.
1446 if (!cpumask_weight(&updated_cpus))
1449 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1452 * Update the numa-cpu lookup table with the new mappings, even for
1453 * offline CPUs. It is best to perform this update from the stop-
1456 stop_machine(update_lookup_table, &updates[0],
1457 cpumask_of(raw_smp_processor_id()));
1459 for (ud = &updates[0]; ud; ud = ud->next) {
1460 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1461 register_cpu_under_node(ud->cpu, ud->new_nid);
1463 dev = get_cpu_device(ud->cpu);
1465 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1466 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1475 static void topology_work_fn(struct work_struct *work)
1477 rebuild_sched_domains();
1479 static DECLARE_WORK(topology_work, topology_work_fn);
1481 static void topology_schedule_update(void)
1483 schedule_work(&topology_work);
1486 static void topology_timer_fn(unsigned long ignored)
1488 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1489 topology_schedule_update();
1490 else if (vphn_enabled) {
1491 if (update_cpu_associativity_changes_mask() > 0)
1492 topology_schedule_update();
1493 reset_topology_timer();
1496 static struct timer_list topology_timer =
1497 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1499 static void reset_topology_timer(void)
1501 topology_timer.data = 0;
1502 topology_timer.expires = jiffies + 60 * HZ;
1503 mod_timer(&topology_timer, topology_timer.expires);
1508 static void stage_topology_update(int core_id)
1510 cpumask_or(&cpu_associativity_changes_mask,
1511 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1512 reset_topology_timer();
1515 static int dt_update_callback(struct notifier_block *nb,
1516 unsigned long action, void *data)
1518 struct of_reconfig_data *update = data;
1519 int rc = NOTIFY_DONE;
1522 case OF_RECONFIG_UPDATE_PROPERTY:
1523 if (!of_prop_cmp(update->dn->type, "cpu") &&
1524 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1526 of_property_read_u32(update->dn, "reg", &core_id);
1527 stage_topology_update(core_id);
1536 static struct notifier_block dt_update_nb = {
1537 .notifier_call = dt_update_callback,
1543 * Start polling for associativity changes.
1545 int start_topology_update(void)
1549 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1550 if (!prrn_enabled) {
1554 rc = of_reconfig_notifier_register(&dt_update_nb);
1557 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1558 lppaca_shared_proc(get_lppaca())) {
1559 if (!vphn_enabled) {
1562 setup_cpu_associativity_change_counters();
1563 init_timer_deferrable(&topology_timer);
1564 reset_topology_timer();
1572 * Disable polling for VPHN associativity changes.
1574 int stop_topology_update(void)
1581 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1583 } else if (vphn_enabled) {
1585 rc = del_timer_sync(&topology_timer);
1591 int prrn_is_enabled(void)
1593 return prrn_enabled;
1596 static int topology_read(struct seq_file *file, void *v)
1598 if (vphn_enabled || prrn_enabled)
1599 seq_puts(file, "on\n");
1601 seq_puts(file, "off\n");
1606 static int topology_open(struct inode *inode, struct file *file)
1608 return single_open(file, topology_read, NULL);
1611 static ssize_t topology_write(struct file *file, const char __user *buf,
1612 size_t count, loff_t *off)
1614 char kbuf[4]; /* "on" or "off" plus null. */
1617 read_len = count < 3 ? count : 3;
1618 if (copy_from_user(kbuf, buf, read_len))
1621 kbuf[read_len] = '\0';
1623 if (!strncmp(kbuf, "on", 2))
1624 start_topology_update();
1625 else if (!strncmp(kbuf, "off", 3))
1626 stop_topology_update();
1633 static const struct file_operations topology_ops = {
1635 .write = topology_write,
1636 .open = topology_open,
1637 .release = single_release
1640 static int topology_update_init(void)
1642 /* Do not poll for changes if disabled at boot */
1643 if (topology_updates_enabled)
1644 start_topology_update();
1646 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1651 device_initcall(topology_update_init);
1652 #endif /* CONFIG_PPC_SPLPAR */