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/memblock.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>
23 #include <linux/pfn.h>
24 #include <linux/cpuset.h>
25 #include <linux/node.h>
26 #include <linux/stop_machine.h>
27 #include <linux/proc_fs.h>
28 #include <linux/seq_file.h>
29 #include <linux/uaccess.h>
30 #include <linux/slab.h>
31 #include <asm/cputhreads.h>
32 #include <asm/sparsemem.h>
35 #include <asm/cputhreads.h>
36 #include <asm/topology.h>
37 #include <asm/firmware.h>
39 #include <asm/hvcall.h>
40 #include <asm/setup.h>
42 #include <asm/drmem.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 */
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 map_cpu_to_node(int cpu, int node)
147 update_numa_cpu_lookup_table(cpu, node);
149 dbg("adding cpu %d to node %d\n", cpu, node);
151 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
152 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
155 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
156 static void unmap_cpu_from_node(unsigned long cpu)
158 int node = numa_cpu_lookup_table[cpu];
160 dbg("removing cpu %lu from node %d\n", cpu, node);
162 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
163 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
165 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
169 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
171 /* must hold reference to node during call */
172 static const __be32 *of_get_associativity(struct device_node *dev)
174 return of_get_property(dev, "ibm,associativity", NULL);
177 int __node_distance(int a, int b)
180 int distance = LOCAL_DISTANCE;
183 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
185 for (i = 0; i < distance_ref_points_depth; i++) {
186 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
189 /* Double the distance for each NUMA level */
195 EXPORT_SYMBOL(__node_distance);
197 static void initialize_distance_lookup_table(int nid,
198 const __be32 *associativity)
205 for (i = 0; i < distance_ref_points_depth; i++) {
208 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
209 distance_lookup_table[nid][i] = of_read_number(entry, 1);
213 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
216 static int associativity_to_nid(const __be32 *associativity)
220 if (min_common_depth == -1)
223 if (of_read_number(associativity, 1) >= min_common_depth)
224 nid = of_read_number(&associativity[min_common_depth], 1);
226 /* POWER4 LPAR uses 0xffff as invalid node */
227 if (nid == 0xffff || nid >= MAX_NUMNODES)
231 of_read_number(associativity, 1) >= distance_ref_points_depth) {
233 * Skip the length field and send start of associativity array
235 initialize_distance_lookup_table(nid, associativity + 1);
242 /* Returns the nid associated with the given device tree node,
243 * or -1 if not found.
245 static int of_node_to_nid_single(struct device_node *device)
250 tmp = of_get_associativity(device);
252 nid = associativity_to_nid(tmp);
256 /* Walk the device tree upwards, looking for an associativity id */
257 int of_node_to_nid(struct device_node *device)
263 nid = of_node_to_nid_single(device);
267 device = of_get_next_parent(device);
273 EXPORT_SYMBOL(of_node_to_nid);
275 static int __init find_min_common_depth(void)
278 struct device_node *root;
280 if (firmware_has_feature(FW_FEATURE_OPAL))
281 root = of_find_node_by_path("/ibm,opal");
283 root = of_find_node_by_path("/rtas");
285 root = of_find_node_by_path("/");
288 * This property is a set of 32-bit integers, each representing
289 * an index into the ibm,associativity nodes.
291 * With form 0 affinity the first integer is for an SMP configuration
292 * (should be all 0's) and the second is for a normal NUMA
293 * configuration. We have only one level of NUMA.
295 * With form 1 affinity the first integer is the most significant
296 * NUMA boundary and the following are progressively less significant
297 * boundaries. There can be more than one level of NUMA.
299 distance_ref_points = of_get_property(root,
300 "ibm,associativity-reference-points",
301 &distance_ref_points_depth);
303 if (!distance_ref_points) {
304 dbg("NUMA: ibm,associativity-reference-points not found.\n");
308 distance_ref_points_depth /= sizeof(int);
310 if (firmware_has_feature(FW_FEATURE_OPAL) ||
311 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
312 dbg("Using form 1 affinity\n");
316 if (form1_affinity) {
317 depth = of_read_number(distance_ref_points, 1);
319 if (distance_ref_points_depth < 2) {
320 printk(KERN_WARNING "NUMA: "
321 "short ibm,associativity-reference-points\n");
325 depth = of_read_number(&distance_ref_points[1], 1);
329 * Warn and cap if the hardware supports more than
330 * MAX_DISTANCE_REF_POINTS domains.
332 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
333 printk(KERN_WARNING "NUMA: distance array capped at "
334 "%d entries\n", MAX_DISTANCE_REF_POINTS);
335 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
346 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
348 struct device_node *memory = NULL;
350 memory = of_find_node_by_type(memory, "memory");
352 panic("numa.c: No memory nodes found!");
354 *n_addr_cells = of_n_addr_cells(memory);
355 *n_size_cells = of_n_size_cells(memory);
359 static unsigned long read_n_cells(int n, const __be32 **buf)
361 unsigned long result = 0;
364 result = (result << 32) | of_read_number(*buf, 1);
370 struct assoc_arrays {
373 const __be32 *arrays;
377 * Retrieve and validate the list of associativity arrays for drconf
378 * memory from the ibm,associativity-lookup-arrays property of the
381 * The layout of the ibm,associativity-lookup-arrays property is a number N
382 * indicating the number of associativity arrays, followed by a number M
383 * indicating the size of each associativity array, followed by a list
384 * of N associativity arrays.
386 static int of_get_assoc_arrays(struct assoc_arrays *aa)
388 struct device_node *memory;
392 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
396 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
397 if (!prop || len < 2 * sizeof(unsigned int)) {
402 aa->n_arrays = of_read_number(prop++, 1);
403 aa->array_sz = of_read_number(prop++, 1);
407 /* Now that we know the number of arrays and size of each array,
408 * revalidate the size of the property read in.
410 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
418 * This is like of_node_to_nid_single() for memory represented in the
419 * ibm,dynamic-reconfiguration-memory node.
421 static int of_drconf_to_nid_single(struct drmem_lmb *lmb)
423 struct assoc_arrays aa = { .arrays = NULL };
425 int nid = default_nid;
428 rc = of_get_assoc_arrays(&aa);
432 if (min_common_depth > 0 && min_common_depth <= aa.array_sz &&
433 !(lmb->flags & DRCONF_MEM_AI_INVALID) &&
434 lmb->aa_index < aa.n_arrays) {
435 index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
436 nid = of_read_number(&aa.arrays[index], 1);
438 if (nid == 0xffff || nid >= MAX_NUMNODES)
442 index = lmb->aa_index * aa.array_sz;
443 initialize_distance_lookup_table(nid,
452 * Figure out to which domain a cpu belongs and stick it there.
453 * Return the id of the domain used.
455 static int numa_setup_cpu(unsigned long lcpu)
458 struct device_node *cpu;
461 * If a valid cpu-to-node mapping is already available, use it
462 * directly instead of querying the firmware, since it represents
463 * the most recent mapping notified to us by the platform (eg: VPHN).
465 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
466 map_cpu_to_node(lcpu, nid);
470 cpu = of_get_cpu_node(lcpu, NULL);
474 if (cpu_present(lcpu))
480 nid = of_node_to_nid_single(cpu);
483 if (nid < 0 || !node_possible(nid))
484 nid = first_online_node;
486 map_cpu_to_node(lcpu, nid);
492 static void verify_cpu_node_mapping(int cpu, int node)
494 int base, sibling, i;
496 /* Verify that all the threads in the core belong to the same node */
497 base = cpu_first_thread_sibling(cpu);
499 for (i = 0; i < threads_per_core; i++) {
502 if (sibling == cpu || cpu_is_offline(sibling))
505 if (cpu_to_node(sibling) != node) {
506 WARN(1, "CPU thread siblings %d and %d don't belong"
507 " to the same node!\n", cpu, sibling);
513 /* Must run before sched domains notifier. */
514 static int ppc_numa_cpu_prepare(unsigned int cpu)
518 nid = numa_setup_cpu(cpu);
519 verify_cpu_node_mapping(cpu, nid);
523 static int ppc_numa_cpu_dead(unsigned int cpu)
525 #ifdef CONFIG_HOTPLUG_CPU
526 unmap_cpu_from_node(cpu);
532 * Check and possibly modify a memory region to enforce the memory limit.
534 * Returns the size the region should have to enforce the memory limit.
535 * This will either be the original value of size, a truncated value,
536 * or zero. If the returned value of size is 0 the region should be
537 * discarded as it lies wholly above the memory limit.
539 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
543 * We use memblock_end_of_DRAM() in here instead of memory_limit because
544 * we've already adjusted it for the limit and it takes care of
545 * having memory holes below the limit. Also, in the case of
546 * iommu_is_off, memory_limit is not set but is implicitly enforced.
549 if (start + size <= memblock_end_of_DRAM())
552 if (start >= memblock_end_of_DRAM())
555 return memblock_end_of_DRAM() - start;
559 * Reads the counter for a given entry in
560 * linux,drconf-usable-memory property
562 static inline int __init read_usm_ranges(const __be32 **usm)
565 * For each lmb in ibm,dynamic-memory a corresponding
566 * entry in linux,drconf-usable-memory property contains
567 * a counter followed by that many (base, size) duple.
568 * read the counter from linux,drconf-usable-memory
570 return read_n_cells(n_mem_size_cells, usm);
574 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
575 * node. This assumes n_mem_{addr,size}_cells have been set.
577 static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
580 unsigned int ranges, is_kexec_kdump = 0;
581 unsigned long base, size, sz;
585 * Skip this block if the reserved bit is set in flags (0x80)
586 * or if the block is not assigned to this partition (0x8)
588 if ((lmb->flags & DRCONF_MEM_RESERVED)
589 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
595 base = lmb->base_addr;
596 size = drmem_lmb_size();
599 if (is_kexec_kdump) {
600 ranges = read_usm_ranges(usm);
601 if (!ranges) /* there are no (base, size) duple */
606 if (is_kexec_kdump) {
607 base = read_n_cells(n_mem_addr_cells, usm);
608 size = read_n_cells(n_mem_size_cells, usm);
611 nid = of_drconf_to_nid_single(lmb);
612 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
614 node_set_online(nid);
615 sz = numa_enforce_memory_limit(base, size);
617 memblock_set_node(base, sz, &memblock.memory, nid);
621 static int __init parse_numa_properties(void)
623 struct device_node *memory;
627 if (numa_enabled == 0) {
628 printk(KERN_WARNING "NUMA disabled by user\n");
632 min_common_depth = find_min_common_depth();
634 if (min_common_depth < 0)
635 return min_common_depth;
637 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
640 * Even though we connect cpus to numa domains later in SMP
641 * init, we need to know the node ids now. This is because
642 * each node to be onlined must have NODE_DATA etc backing it.
644 for_each_present_cpu(i) {
645 struct device_node *cpu;
648 cpu = of_get_cpu_node(i, NULL);
650 nid = of_node_to_nid_single(cpu);
654 * Don't fall back to default_nid yet -- we will plug
655 * cpus into nodes once the memory scan has discovered
660 node_set_online(nid);
663 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
665 for_each_node_by_type(memory, "memory") {
670 const __be32 *memcell_buf;
673 memcell_buf = of_get_property(memory,
674 "linux,usable-memory", &len);
675 if (!memcell_buf || len <= 0)
676 memcell_buf = of_get_property(memory, "reg", &len);
677 if (!memcell_buf || len <= 0)
681 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
683 /* these are order-sensitive, and modify the buffer pointer */
684 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
685 size = read_n_cells(n_mem_size_cells, &memcell_buf);
688 * Assumption: either all memory nodes or none will
689 * have associativity properties. If none, then
690 * everything goes to default_nid.
692 nid = of_node_to_nid_single(memory);
696 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
697 node_set_online(nid);
699 size = numa_enforce_memory_limit(start, size);
701 memblock_set_node(start, size, &memblock.memory, nid);
708 * Now do the same thing for each MEMBLOCK listed in the
709 * ibm,dynamic-memory property in the
710 * ibm,dynamic-reconfiguration-memory node.
712 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
714 walk_drmem_lmbs(memory, numa_setup_drmem_lmb);
721 static void __init setup_nonnuma(void)
723 unsigned long top_of_ram = memblock_end_of_DRAM();
724 unsigned long total_ram = memblock_phys_mem_size();
725 unsigned long start_pfn, end_pfn;
726 unsigned int nid = 0;
727 struct memblock_region *reg;
729 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
730 top_of_ram, total_ram);
731 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
732 (top_of_ram - total_ram) >> 20);
734 for_each_memblock(memory, reg) {
735 start_pfn = memblock_region_memory_base_pfn(reg);
736 end_pfn = memblock_region_memory_end_pfn(reg);
738 fake_numa_create_new_node(end_pfn, &nid);
739 memblock_set_node(PFN_PHYS(start_pfn),
740 PFN_PHYS(end_pfn - start_pfn),
741 &memblock.memory, nid);
742 node_set_online(nid);
746 void __init dump_numa_cpu_topology(void)
749 unsigned int cpu, count;
751 if (min_common_depth == -1 || !numa_enabled)
754 for_each_online_node(node) {
755 pr_info("Node %d CPUs:", node);
759 * If we used a CPU iterator here we would miss printing
760 * the holes in the cpumap.
762 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
763 if (cpumask_test_cpu(cpu,
764 node_to_cpumask_map[node])) {
770 pr_cont("-%u", cpu - 1);
776 pr_cont("-%u", nr_cpu_ids - 1);
781 /* Initialize NODE_DATA for a node on the local memory */
782 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
784 u64 spanned_pages = end_pfn - start_pfn;
785 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
790 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
793 /* report and initialize */
794 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
795 nd_pa, nd_pa + nd_size - 1);
796 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
798 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
801 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
802 NODE_DATA(nid)->node_id = nid;
803 NODE_DATA(nid)->node_start_pfn = start_pfn;
804 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
807 static void __init find_possible_nodes(void)
809 struct device_node *rtas;
812 if (min_common_depth <= 0)
815 rtas = of_find_node_by_path("/rtas");
819 if (of_property_read_u32_index(rtas,
820 "ibm,max-associativity-domains",
821 min_common_depth, &numnodes))
824 for (i = 0; i < numnodes; i++) {
825 if (!node_possible(i))
826 node_set(i, node_possible_map);
833 void __init mem_topology_setup(void)
837 if (parse_numa_properties())
841 * Modify the set of possible NUMA nodes to reflect information
842 * available about the set of online nodes, and the set of nodes
843 * that we expect to make use of for this platform's affinity
846 nodes_and(node_possible_map, node_possible_map, node_online_map);
848 find_possible_nodes();
850 setup_node_to_cpumask_map();
852 reset_numa_cpu_lookup_table();
854 for_each_present_cpu(cpu)
858 void __init initmem_init(void)
862 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
863 max_pfn = max_low_pfn;
867 for_each_online_node(nid) {
868 unsigned long start_pfn, end_pfn;
870 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
871 setup_node_data(nid, start_pfn, end_pfn);
872 sparse_memory_present_with_active_regions(nid);
878 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
879 * even before we online them, so that we can use cpu_to_{node,mem}
880 * early in boot, cf. smp_prepare_cpus().
881 * _nocalls() + manual invocation is used because cpuhp is not yet
882 * initialized for the boot CPU.
884 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
885 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
888 static int __init early_numa(char *p)
893 if (strstr(p, "off"))
896 if (strstr(p, "debug"))
899 p = strstr(p, "fake=");
901 cmdline = p + strlen("fake=");
905 early_param("numa", early_numa);
907 static bool topology_updates_enabled = true;
909 static int __init early_topology_updates(char *p)
914 if (!strcmp(p, "off")) {
915 pr_info("Disabling topology updates\n");
916 topology_updates_enabled = false;
921 early_param("topology_updates", early_topology_updates);
923 #ifdef CONFIG_MEMORY_HOTPLUG
925 * Find the node associated with a hot added memory section for
926 * memory represented in the device tree by the property
927 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
929 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
931 struct drmem_lmb *lmb;
932 unsigned long lmb_size;
935 lmb_size = drmem_lmb_size();
937 for_each_drmem_lmb(lmb) {
938 /* skip this block if it is reserved or not assigned to
940 if ((lmb->flags & DRCONF_MEM_RESERVED)
941 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
944 if ((scn_addr < lmb->base_addr)
945 || (scn_addr >= (lmb->base_addr + lmb_size)))
948 nid = of_drconf_to_nid_single(lmb);
956 * Find the node associated with a hot added memory section for memory
957 * represented in the device tree as a node (i.e. memory@XXXX) for
960 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
962 struct device_node *memory;
965 for_each_node_by_type(memory, "memory") {
966 unsigned long start, size;
968 const __be32 *memcell_buf;
971 memcell_buf = of_get_property(memory, "reg", &len);
972 if (!memcell_buf || len <= 0)
976 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
979 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
980 size = read_n_cells(n_mem_size_cells, &memcell_buf);
982 if ((scn_addr < start) || (scn_addr >= (start + size)))
985 nid = of_node_to_nid_single(memory);
999 * Find the node associated with a hot added memory section. Section
1000 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1001 * sections are fully contained within a single MEMBLOCK.
1003 int hot_add_scn_to_nid(unsigned long scn_addr)
1005 struct device_node *memory = NULL;
1008 if (!numa_enabled || (min_common_depth < 0))
1009 return first_online_node;
1011 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1013 nid = hot_add_drconf_scn_to_nid(scn_addr);
1014 of_node_put(memory);
1016 nid = hot_add_node_scn_to_nid(scn_addr);
1019 if (nid < 0 || !node_possible(nid))
1020 nid = first_online_node;
1025 static u64 hot_add_drconf_memory_max(void)
1027 struct device_node *memory = NULL;
1028 struct device_node *dn = NULL;
1029 const __be64 *lrdr = NULL;
1031 dn = of_find_node_by_path("/rtas");
1033 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1036 return be64_to_cpup(lrdr);
1039 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1041 of_node_put(memory);
1042 return drmem_lmb_memory_max();
1048 * memory_hotplug_max - return max address of memory that may be added
1050 * This is currently only used on systems that support drconfig memory
1053 u64 memory_hotplug_max(void)
1055 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1057 #endif /* CONFIG_MEMORY_HOTPLUG */
1059 /* Virtual Processor Home Node (VPHN) support */
1060 #ifdef CONFIG_PPC_SPLPAR
1064 struct topology_update_data {
1065 struct topology_update_data *next;
1071 #define TOPOLOGY_DEF_TIMER_SECS 60
1073 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1074 static cpumask_t cpu_associativity_changes_mask;
1075 static int vphn_enabled;
1076 static int prrn_enabled;
1077 static void reset_topology_timer(void);
1078 static int topology_timer_secs = 1;
1079 static int topology_inited;
1082 * Change polling interval for associativity changes.
1084 int timed_topology_update(int nsecs)
1088 topology_timer_secs = nsecs;
1090 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1092 reset_topology_timer();
1099 * Store the current values of the associativity change counters in the
1102 static void setup_cpu_associativity_change_counters(void)
1106 /* The VPHN feature supports a maximum of 8 reference points */
1107 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1109 for_each_possible_cpu(cpu) {
1111 u8 *counts = vphn_cpu_change_counts[cpu];
1112 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1114 for (i = 0; i < distance_ref_points_depth; i++)
1115 counts[i] = hypervisor_counts[i];
1120 * The hypervisor maintains a set of 8 associativity change counters in
1121 * the VPA of each cpu that correspond to the associativity levels in the
1122 * ibm,associativity-reference-points property. When an associativity
1123 * level changes, the corresponding counter is incremented.
1125 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1126 * node associativity levels have changed.
1128 * Returns the number of cpus with unhandled associativity changes.
1130 static int update_cpu_associativity_changes_mask(void)
1133 cpumask_t *changes = &cpu_associativity_changes_mask;
1135 for_each_possible_cpu(cpu) {
1137 u8 *counts = vphn_cpu_change_counts[cpu];
1138 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1140 for (i = 0; i < distance_ref_points_depth; i++) {
1141 if (hypervisor_counts[i] != counts[i]) {
1142 counts[i] = hypervisor_counts[i];
1147 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1148 cpu = cpu_last_thread_sibling(cpu);
1152 return cpumask_weight(changes);
1156 * Retrieve the new associativity information for a virtual processor's
1159 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1162 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1164 int hwcpu = get_hard_smp_processor_id(cpu);
1166 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1167 vphn_unpack_associativity(retbuf, associativity);
1172 static long vphn_get_associativity(unsigned long cpu,
1173 __be32 *associativity)
1177 rc = hcall_vphn(cpu, associativity);
1182 "VPHN is not supported. Disabling polling...\n");
1183 stop_topology_update();
1187 "hcall_vphn() experienced a hardware fault "
1188 "preventing VPHN. Disabling polling...\n");
1189 stop_topology_update();
1192 dbg("VPHN hcall succeeded. Reset polling...\n");
1193 timed_topology_update(0);
1200 int find_and_online_cpu_nid(int cpu)
1202 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1205 /* Use associativity from first thread for all siblings */
1206 if (vphn_get_associativity(cpu, associativity))
1207 return cpu_to_node(cpu);
1209 new_nid = associativity_to_nid(associativity);
1210 if (new_nid < 0 || !node_possible(new_nid))
1211 new_nid = first_online_node;
1213 if (NODE_DATA(new_nid) == NULL) {
1214 #ifdef CONFIG_MEMORY_HOTPLUG
1216 * Need to ensure that NODE_DATA is initialized for a node from
1217 * available memory (see memblock_alloc_try_nid). If unable to
1218 * init the node, then default to nearest node that has memory
1219 * installed. Skip onlining a node if the subsystems are not
1222 if (!topology_inited || try_online_node(new_nid))
1223 new_nid = first_online_node;
1226 * Default to using the nearest node that has memory installed.
1227 * Otherwise, it would be necessary to patch the kernel MM code
1228 * to deal with more memoryless-node error conditions.
1230 new_nid = first_online_node;
1234 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1240 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1241 * characteristics change. This function doesn't perform any locking and is
1242 * only safe to call from stop_machine().
1244 static int update_cpu_topology(void *data)
1246 struct topology_update_data *update;
1252 cpu = smp_processor_id();
1254 for (update = data; update; update = update->next) {
1255 int new_nid = update->new_nid;
1256 if (cpu != update->cpu)
1259 unmap_cpu_from_node(cpu);
1260 map_cpu_to_node(cpu, new_nid);
1261 set_cpu_numa_node(cpu, new_nid);
1262 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1269 static int update_lookup_table(void *data)
1271 struct topology_update_data *update;
1277 * Upon topology update, the numa-cpu lookup table needs to be updated
1278 * for all threads in the core, including offline CPUs, to ensure that
1279 * future hotplug operations respect the cpu-to-node associativity
1282 for (update = data; update; update = update->next) {
1285 nid = update->new_nid;
1286 base = cpu_first_thread_sibling(update->cpu);
1288 for (j = 0; j < threads_per_core; j++) {
1289 update_numa_cpu_lookup_table(base + j, nid);
1297 * Update the node maps and sysfs entries for each cpu whose home node
1298 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1300 * cpus_locked says whether we already hold cpu_hotplug_lock.
1302 int numa_update_cpu_topology(bool cpus_locked)
1304 unsigned int cpu, sibling, changed = 0;
1305 struct topology_update_data *updates, *ud;
1306 cpumask_t updated_cpus;
1308 int weight, new_nid, i = 0;
1310 if (!prrn_enabled && !vphn_enabled && topology_inited)
1313 weight = cpumask_weight(&cpu_associativity_changes_mask);
1317 updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL);
1321 cpumask_clear(&updated_cpus);
1323 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1325 * If siblings aren't flagged for changes, updates list
1326 * will be too short. Skip on this update and set for next
1329 if (!cpumask_subset(cpu_sibling_mask(cpu),
1330 &cpu_associativity_changes_mask)) {
1331 pr_info("Sibling bits not set for associativity "
1332 "change, cpu%d\n", cpu);
1333 cpumask_or(&cpu_associativity_changes_mask,
1334 &cpu_associativity_changes_mask,
1335 cpu_sibling_mask(cpu));
1336 cpu = cpu_last_thread_sibling(cpu);
1340 new_nid = find_and_online_cpu_nid(cpu);
1342 if (new_nid == numa_cpu_lookup_table[cpu]) {
1343 cpumask_andnot(&cpu_associativity_changes_mask,
1344 &cpu_associativity_changes_mask,
1345 cpu_sibling_mask(cpu));
1346 dbg("Assoc chg gives same node %d for cpu%d\n",
1348 cpu = cpu_last_thread_sibling(cpu);
1352 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1354 ud->next = &updates[i];
1356 ud->new_nid = new_nid;
1357 ud->old_nid = numa_cpu_lookup_table[sibling];
1358 cpumask_set_cpu(sibling, &updated_cpus);
1360 cpu = cpu_last_thread_sibling(cpu);
1364 * Prevent processing of 'updates' from overflowing array
1365 * where last entry filled in a 'next' pointer.
1368 updates[i-1].next = NULL;
1370 pr_debug("Topology update for the following CPUs:\n");
1371 if (cpumask_weight(&updated_cpus)) {
1372 for (ud = &updates[0]; ud; ud = ud->next) {
1373 pr_debug("cpu %d moving from node %d "
1375 ud->old_nid, ud->new_nid);
1380 * In cases where we have nothing to update (because the updates list
1381 * is too short or because the new topology is same as the old one),
1382 * skip invoking update_cpu_topology() via stop-machine(). This is
1383 * necessary (and not just a fast-path optimization) since stop-machine
1384 * can end up electing a random CPU to run update_cpu_topology(), and
1385 * thus trick us into setting up incorrect cpu-node mappings (since
1386 * 'updates' is kzalloc()'ed).
1388 * And for the similar reason, we will skip all the following updating.
1390 if (!cpumask_weight(&updated_cpus))
1394 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1397 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1400 * Update the numa-cpu lookup table with the new mappings, even for
1401 * offline CPUs. It is best to perform this update from the stop-
1405 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1406 cpumask_of(raw_smp_processor_id()));
1408 stop_machine(update_lookup_table, &updates[0],
1409 cpumask_of(raw_smp_processor_id()));
1411 for (ud = &updates[0]; ud; ud = ud->next) {
1412 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1413 register_cpu_under_node(ud->cpu, ud->new_nid);
1415 dev = get_cpu_device(ud->cpu);
1417 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1418 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1427 int arch_update_cpu_topology(void)
1429 return numa_update_cpu_topology(true);
1432 static void topology_work_fn(struct work_struct *work)
1434 rebuild_sched_domains();
1436 static DECLARE_WORK(topology_work, topology_work_fn);
1438 static void topology_schedule_update(void)
1440 schedule_work(&topology_work);
1443 static void topology_timer_fn(struct timer_list *unused)
1445 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1446 topology_schedule_update();
1447 else if (vphn_enabled) {
1448 if (update_cpu_associativity_changes_mask() > 0)
1449 topology_schedule_update();
1450 reset_topology_timer();
1453 static struct timer_list topology_timer;
1455 static void reset_topology_timer(void)
1458 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1463 static void stage_topology_update(int core_id)
1465 cpumask_or(&cpu_associativity_changes_mask,
1466 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1467 reset_topology_timer();
1470 static int dt_update_callback(struct notifier_block *nb,
1471 unsigned long action, void *data)
1473 struct of_reconfig_data *update = data;
1474 int rc = NOTIFY_DONE;
1477 case OF_RECONFIG_UPDATE_PROPERTY:
1478 if (!of_prop_cmp(update->dn->type, "cpu") &&
1479 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1481 of_property_read_u32(update->dn, "reg", &core_id);
1482 stage_topology_update(core_id);
1491 static struct notifier_block dt_update_nb = {
1492 .notifier_call = dt_update_callback,
1498 * Start polling for associativity changes.
1500 int start_topology_update(void)
1504 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1505 if (!prrn_enabled) {
1508 rc = of_reconfig_notifier_register(&dt_update_nb);
1512 if (firmware_has_feature(FW_FEATURE_VPHN) &&
1513 lppaca_shared_proc(get_lppaca())) {
1514 if (!vphn_enabled) {
1516 setup_cpu_associativity_change_counters();
1517 timer_setup(&topology_timer, topology_timer_fn,
1519 reset_topology_timer();
1523 pr_info("Starting topology update%s%s\n",
1524 (prrn_enabled ? " prrn_enabled" : ""),
1525 (vphn_enabled ? " vphn_enabled" : ""));
1531 * Disable polling for VPHN associativity changes.
1533 int stop_topology_update(void)
1540 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1545 rc = del_timer_sync(&topology_timer);
1548 pr_info("Stopping topology update\n");
1553 int prrn_is_enabled(void)
1555 return prrn_enabled;
1558 void __init shared_proc_topology_init(void)
1560 if (lppaca_shared_proc(get_lppaca())) {
1561 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1563 numa_update_cpu_topology(false);
1567 static int topology_read(struct seq_file *file, void *v)
1569 if (vphn_enabled || prrn_enabled)
1570 seq_puts(file, "on\n");
1572 seq_puts(file, "off\n");
1577 static int topology_open(struct inode *inode, struct file *file)
1579 return single_open(file, topology_read, NULL);
1582 static ssize_t topology_write(struct file *file, const char __user *buf,
1583 size_t count, loff_t *off)
1585 char kbuf[4]; /* "on" or "off" plus null. */
1588 read_len = count < 3 ? count : 3;
1589 if (copy_from_user(kbuf, buf, read_len))
1592 kbuf[read_len] = '\0';
1594 if (!strncmp(kbuf, "on", 2))
1595 start_topology_update();
1596 else if (!strncmp(kbuf, "off", 3))
1597 stop_topology_update();
1604 static const struct file_operations topology_ops = {
1606 .write = topology_write,
1607 .open = topology_open,
1608 .release = single_release
1611 static int topology_update_init(void)
1613 /* Do not poll for changes if disabled at boot */
1614 if (topology_updates_enabled)
1615 start_topology_update();
1618 topology_schedule_update();
1620 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1623 topology_inited = 1;
1626 device_initcall(topology_update_init);
1627 #endif /* CONFIG_PPC_SPLPAR */