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/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/topology.h>
36 #include <asm/firmware.h>
38 #include <asm/hvcall.h>
39 #include <asm/setup.h>
41 #include <asm/drmem.h>
43 static int numa_enabled = 1;
45 static char *cmdline __initdata;
47 static int numa_debug;
48 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
50 int numa_cpu_lookup_table[NR_CPUS];
51 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
52 struct pglist_data *node_data[MAX_NUMNODES];
54 EXPORT_SYMBOL(numa_cpu_lookup_table);
55 EXPORT_SYMBOL(node_to_cpumask_map);
56 EXPORT_SYMBOL(node_data);
58 static int min_common_depth;
59 static int n_mem_addr_cells, n_mem_size_cells;
60 static int form1_affinity;
62 #define MAX_DISTANCE_REF_POINTS 4
63 static int distance_ref_points_depth;
64 static const __be32 *distance_ref_points;
65 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
68 * Allocate node_to_cpumask_map based on number of available nodes
69 * Requires node_possible_map to be valid.
71 * Note: cpumask_of_node() is not valid until after this is done.
73 static void __init setup_node_to_cpumask_map(void)
77 /* setup nr_node_ids if not done yet */
78 if (nr_node_ids == MAX_NUMNODES)
81 /* allocate the map */
83 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
85 /* cpumask_of_node() will now work */
86 dbg("Node to cpumask map for %u nodes\n", nr_node_ids);
89 static int __init fake_numa_create_new_node(unsigned long end_pfn,
92 unsigned long long mem;
94 static unsigned int fake_nid;
95 static unsigned long long curr_boundary;
98 * Modify node id, iff we started creating NUMA nodes
99 * We want to continue from where we left of the last time
104 * In case there are no more arguments to parse, the
105 * node_id should be the same as the last fake node id
106 * (we've handled this above).
111 mem = memparse(p, &p);
115 if (mem < curr_boundary)
120 if ((end_pfn << PAGE_SHIFT) > mem) {
122 * Skip commas and spaces
124 while (*p == ',' || *p == ' ' || *p == '\t')
130 dbg("created new fake_node with id %d\n", fake_nid);
136 static void reset_numa_cpu_lookup_table(void)
140 for_each_possible_cpu(cpu)
141 numa_cpu_lookup_table[cpu] = -1;
144 static void map_cpu_to_node(int cpu, int node)
146 update_numa_cpu_lookup_table(cpu, node);
148 dbg("adding cpu %d to node %d\n", cpu, node);
150 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
151 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
154 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
155 static void unmap_cpu_from_node(unsigned long cpu)
157 int node = numa_cpu_lookup_table[cpu];
159 dbg("removing cpu %lu from node %d\n", cpu, node);
161 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
162 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
164 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
168 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
170 int cpu_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
176 for (i = 0; i < distance_ref_points_depth; i++) {
177 index = be32_to_cpu(distance_ref_points[i]);
178 if (cpu1_assoc[index] == cpu2_assoc[index])
186 /* must hold reference to node during call */
187 static const __be32 *of_get_associativity(struct device_node *dev)
189 return of_get_property(dev, "ibm,associativity", NULL);
192 int __node_distance(int a, int b)
195 int distance = LOCAL_DISTANCE;
198 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
200 for (i = 0; i < distance_ref_points_depth; i++) {
201 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
204 /* Double the distance for each NUMA level */
210 EXPORT_SYMBOL(__node_distance);
212 static void initialize_distance_lookup_table(int nid,
213 const __be32 *associativity)
220 for (i = 0; i < distance_ref_points_depth; i++) {
223 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
224 distance_lookup_table[nid][i] = of_read_number(entry, 1);
228 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
231 static int associativity_to_nid(const __be32 *associativity)
233 int nid = NUMA_NO_NODE;
238 if (of_read_number(associativity, 1) >= min_common_depth)
239 nid = of_read_number(&associativity[min_common_depth], 1);
241 /* POWER4 LPAR uses 0xffff as invalid node */
242 if (nid == 0xffff || nid >= MAX_NUMNODES)
246 of_read_number(associativity, 1) >= distance_ref_points_depth) {
248 * Skip the length field and send start of associativity array
250 initialize_distance_lookup_table(nid, associativity + 1);
257 /* Returns the nid associated with the given device tree node,
258 * or -1 if not found.
260 static int of_node_to_nid_single(struct device_node *device)
262 int nid = NUMA_NO_NODE;
265 tmp = of_get_associativity(device);
267 nid = associativity_to_nid(tmp);
271 /* Walk the device tree upwards, looking for an associativity id */
272 int of_node_to_nid(struct device_node *device)
274 int nid = NUMA_NO_NODE;
278 nid = of_node_to_nid_single(device);
282 device = of_get_next_parent(device);
288 EXPORT_SYMBOL(of_node_to_nid);
290 static int __init find_min_common_depth(void)
293 struct device_node *root;
295 if (firmware_has_feature(FW_FEATURE_OPAL))
296 root = of_find_node_by_path("/ibm,opal");
298 root = of_find_node_by_path("/rtas");
300 root = of_find_node_by_path("/");
303 * This property is a set of 32-bit integers, each representing
304 * an index into the ibm,associativity nodes.
306 * With form 0 affinity the first integer is for an SMP configuration
307 * (should be all 0's) and the second is for a normal NUMA
308 * configuration. We have only one level of NUMA.
310 * With form 1 affinity the first integer is the most significant
311 * NUMA boundary and the following are progressively less significant
312 * boundaries. There can be more than one level of NUMA.
314 distance_ref_points = of_get_property(root,
315 "ibm,associativity-reference-points",
316 &distance_ref_points_depth);
318 if (!distance_ref_points) {
319 dbg("NUMA: ibm,associativity-reference-points not found.\n");
323 distance_ref_points_depth /= sizeof(int);
325 if (firmware_has_feature(FW_FEATURE_OPAL) ||
326 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
327 dbg("Using form 1 affinity\n");
331 if (form1_affinity) {
332 depth = of_read_number(distance_ref_points, 1);
334 if (distance_ref_points_depth < 2) {
335 printk(KERN_WARNING "NUMA: "
336 "short ibm,associativity-reference-points\n");
340 depth = of_read_number(&distance_ref_points[1], 1);
344 * Warn and cap if the hardware supports more than
345 * MAX_DISTANCE_REF_POINTS domains.
347 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
348 printk(KERN_WARNING "NUMA: distance array capped at "
349 "%d entries\n", MAX_DISTANCE_REF_POINTS);
350 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
361 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
363 struct device_node *memory = NULL;
365 memory = of_find_node_by_type(memory, "memory");
367 panic("numa.c: No memory nodes found!");
369 *n_addr_cells = of_n_addr_cells(memory);
370 *n_size_cells = of_n_size_cells(memory);
374 static unsigned long read_n_cells(int n, const __be32 **buf)
376 unsigned long result = 0;
379 result = (result << 32) | of_read_number(*buf, 1);
385 struct assoc_arrays {
388 const __be32 *arrays;
392 * Retrieve and validate the list of associativity arrays for drconf
393 * memory from the ibm,associativity-lookup-arrays property of the
396 * The layout of the ibm,associativity-lookup-arrays property is a number N
397 * indicating the number of associativity arrays, followed by a number M
398 * indicating the size of each associativity array, followed by a list
399 * of N associativity arrays.
401 static int of_get_assoc_arrays(struct assoc_arrays *aa)
403 struct device_node *memory;
407 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
411 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
412 if (!prop || len < 2 * sizeof(unsigned int)) {
417 aa->n_arrays = of_read_number(prop++, 1);
418 aa->array_sz = of_read_number(prop++, 1);
422 /* Now that we know the number of arrays and size of each array,
423 * revalidate the size of the property read in.
425 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
433 * This is like of_node_to_nid_single() for memory represented in the
434 * ibm,dynamic-reconfiguration-memory node.
436 static int of_drconf_to_nid_single(struct drmem_lmb *lmb)
438 struct assoc_arrays aa = { .arrays = NULL };
439 int default_nid = NUMA_NO_NODE;
440 int nid = default_nid;
443 if ((min_common_depth < 0) || !numa_enabled)
446 rc = of_get_assoc_arrays(&aa);
450 if (min_common_depth <= aa.array_sz &&
451 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
452 index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
453 nid = of_read_number(&aa.arrays[index], 1);
455 if (nid == 0xffff || nid >= MAX_NUMNODES)
459 index = lmb->aa_index * aa.array_sz;
460 initialize_distance_lookup_table(nid,
469 * Figure out to which domain a cpu belongs and stick it there.
470 * Return the id of the domain used.
472 static int numa_setup_cpu(unsigned long lcpu)
474 int nid = NUMA_NO_NODE;
475 struct device_node *cpu;
478 * If a valid cpu-to-node mapping is already available, use it
479 * directly instead of querying the firmware, since it represents
480 * the most recent mapping notified to us by the platform (eg: VPHN).
482 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
483 map_cpu_to_node(lcpu, nid);
487 cpu = of_get_cpu_node(lcpu, NULL);
491 if (cpu_present(lcpu))
497 nid = of_node_to_nid_single(cpu);
500 if (nid < 0 || !node_possible(nid))
501 nid = first_online_node;
503 map_cpu_to_node(lcpu, nid);
509 static void verify_cpu_node_mapping(int cpu, int node)
511 int base, sibling, i;
513 /* Verify that all the threads in the core belong to the same node */
514 base = cpu_first_thread_sibling(cpu);
516 for (i = 0; i < threads_per_core; i++) {
519 if (sibling == cpu || cpu_is_offline(sibling))
522 if (cpu_to_node(sibling) != node) {
523 WARN(1, "CPU thread siblings %d and %d don't belong"
524 " to the same node!\n", cpu, sibling);
530 /* Must run before sched domains notifier. */
531 static int ppc_numa_cpu_prepare(unsigned int cpu)
535 nid = numa_setup_cpu(cpu);
536 verify_cpu_node_mapping(cpu, nid);
540 static int ppc_numa_cpu_dead(unsigned int cpu)
542 #ifdef CONFIG_HOTPLUG_CPU
543 unmap_cpu_from_node(cpu);
549 * Check and possibly modify a memory region to enforce the memory limit.
551 * Returns the size the region should have to enforce the memory limit.
552 * This will either be the original value of size, a truncated value,
553 * or zero. If the returned value of size is 0 the region should be
554 * discarded as it lies wholly above the memory limit.
556 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
560 * We use memblock_end_of_DRAM() in here instead of memory_limit because
561 * we've already adjusted it for the limit and it takes care of
562 * having memory holes below the limit. Also, in the case of
563 * iommu_is_off, memory_limit is not set but is implicitly enforced.
566 if (start + size <= memblock_end_of_DRAM())
569 if (start >= memblock_end_of_DRAM())
572 return memblock_end_of_DRAM() - start;
576 * Reads the counter for a given entry in
577 * linux,drconf-usable-memory property
579 static inline int __init read_usm_ranges(const __be32 **usm)
582 * For each lmb in ibm,dynamic-memory a corresponding
583 * entry in linux,drconf-usable-memory property contains
584 * a counter followed by that many (base, size) duple.
585 * read the counter from linux,drconf-usable-memory
587 return read_n_cells(n_mem_size_cells, usm);
591 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
592 * node. This assumes n_mem_{addr,size}_cells have been set.
594 static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
597 unsigned int ranges, is_kexec_kdump = 0;
598 unsigned long base, size, sz;
602 * Skip this block if the reserved bit is set in flags (0x80)
603 * or if the block is not assigned to this partition (0x8)
605 if ((lmb->flags & DRCONF_MEM_RESERVED)
606 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
612 base = lmb->base_addr;
613 size = drmem_lmb_size();
616 if (is_kexec_kdump) {
617 ranges = read_usm_ranges(usm);
618 if (!ranges) /* there are no (base, size) duple */
623 if (is_kexec_kdump) {
624 base = read_n_cells(n_mem_addr_cells, usm);
625 size = read_n_cells(n_mem_size_cells, usm);
628 nid = of_drconf_to_nid_single(lmb);
629 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
631 node_set_online(nid);
632 sz = numa_enforce_memory_limit(base, size);
634 memblock_set_node(base, sz, &memblock.memory, nid);
638 static int __init parse_numa_properties(void)
640 struct device_node *memory;
644 if (numa_enabled == 0) {
645 printk(KERN_WARNING "NUMA disabled by user\n");
649 min_common_depth = find_min_common_depth();
651 if (min_common_depth < 0) {
653 * if we fail to parse min_common_depth from device tree
654 * mark the numa disabled, boot with numa disabled.
656 numa_enabled = false;
657 return min_common_depth;
660 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
663 * Even though we connect cpus to numa domains later in SMP
664 * init, we need to know the node ids now. This is because
665 * each node to be onlined must have NODE_DATA etc backing it.
667 for_each_present_cpu(i) {
668 struct device_node *cpu;
671 cpu = of_get_cpu_node(i, NULL);
673 nid = of_node_to_nid_single(cpu);
677 * Don't fall back to default_nid yet -- we will plug
678 * cpus into nodes once the memory scan has discovered
683 node_set_online(nid);
686 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
688 for_each_node_by_type(memory, "memory") {
693 const __be32 *memcell_buf;
696 memcell_buf = of_get_property(memory,
697 "linux,usable-memory", &len);
698 if (!memcell_buf || len <= 0)
699 memcell_buf = of_get_property(memory, "reg", &len);
700 if (!memcell_buf || len <= 0)
704 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
706 /* these are order-sensitive, and modify the buffer pointer */
707 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
708 size = read_n_cells(n_mem_size_cells, &memcell_buf);
711 * Assumption: either all memory nodes or none will
712 * have associativity properties. If none, then
713 * everything goes to default_nid.
715 nid = of_node_to_nid_single(memory);
719 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
720 node_set_online(nid);
722 size = numa_enforce_memory_limit(start, size);
724 memblock_set_node(start, size, &memblock.memory, nid);
731 * Now do the same thing for each MEMBLOCK listed in the
732 * ibm,dynamic-memory property in the
733 * ibm,dynamic-reconfiguration-memory node.
735 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
737 walk_drmem_lmbs(memory, numa_setup_drmem_lmb);
744 static void __init setup_nonnuma(void)
746 unsigned long top_of_ram = memblock_end_of_DRAM();
747 unsigned long total_ram = memblock_phys_mem_size();
748 unsigned long start_pfn, end_pfn;
749 unsigned int nid = 0;
750 struct memblock_region *reg;
752 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
753 top_of_ram, total_ram);
754 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
755 (top_of_ram - total_ram) >> 20);
757 for_each_memblock(memory, reg) {
758 start_pfn = memblock_region_memory_base_pfn(reg);
759 end_pfn = memblock_region_memory_end_pfn(reg);
761 fake_numa_create_new_node(end_pfn, &nid);
762 memblock_set_node(PFN_PHYS(start_pfn),
763 PFN_PHYS(end_pfn - start_pfn),
764 &memblock.memory, nid);
765 node_set_online(nid);
769 void __init dump_numa_cpu_topology(void)
772 unsigned int cpu, count;
777 for_each_online_node(node) {
778 pr_info("Node %d CPUs:", node);
782 * If we used a CPU iterator here we would miss printing
783 * the holes in the cpumap.
785 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
786 if (cpumask_test_cpu(cpu,
787 node_to_cpumask_map[node])) {
793 pr_cont("-%u", cpu - 1);
799 pr_cont("-%u", nr_cpu_ids - 1);
804 /* Initialize NODE_DATA for a node on the local memory */
805 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
807 u64 spanned_pages = end_pfn - start_pfn;
808 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
813 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
815 panic("Cannot allocate %zu bytes for node %d data\n",
820 /* report and initialize */
821 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
822 nd_pa, nd_pa + nd_size - 1);
823 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
825 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
828 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
829 NODE_DATA(nid)->node_id = nid;
830 NODE_DATA(nid)->node_start_pfn = start_pfn;
831 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
834 static void __init find_possible_nodes(void)
836 struct device_node *rtas;
842 rtas = of_find_node_by_path("/rtas");
846 if (of_property_read_u32_index(rtas,
847 "ibm,max-associativity-domains",
848 min_common_depth, &numnodes))
851 for (i = 0; i < numnodes; i++) {
852 if (!node_possible(i))
853 node_set(i, node_possible_map);
860 void __init mem_topology_setup(void)
864 if (parse_numa_properties())
868 * Modify the set of possible NUMA nodes to reflect information
869 * available about the set of online nodes, and the set of nodes
870 * that we expect to make use of for this platform's affinity
873 nodes_and(node_possible_map, node_possible_map, node_online_map);
875 find_possible_nodes();
877 setup_node_to_cpumask_map();
879 reset_numa_cpu_lookup_table();
881 for_each_present_cpu(cpu)
885 void __init initmem_init(void)
889 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
890 max_pfn = max_low_pfn;
894 for_each_online_node(nid) {
895 unsigned long start_pfn, end_pfn;
897 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
898 setup_node_data(nid, start_pfn, end_pfn);
899 sparse_memory_present_with_active_regions(nid);
905 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
906 * even before we online them, so that we can use cpu_to_{node,mem}
907 * early in boot, cf. smp_prepare_cpus().
908 * _nocalls() + manual invocation is used because cpuhp is not yet
909 * initialized for the boot CPU.
911 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
912 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
915 static int __init early_numa(char *p)
920 if (strstr(p, "off"))
923 if (strstr(p, "debug"))
926 p = strstr(p, "fake=");
928 cmdline = p + strlen("fake=");
932 early_param("numa", early_numa);
935 * The platform can inform us through one of several mechanisms
936 * (post-migration device tree updates, PRRN or VPHN) that the NUMA
937 * assignment of a resource has changed. This controls whether we act
938 * on that. Disabled by default.
940 static bool topology_updates_enabled;
942 static int __init early_topology_updates(char *p)
947 if (!strcmp(p, "on")) {
948 pr_warn("Caution: enabling topology updates\n");
949 topology_updates_enabled = true;
954 early_param("topology_updates", early_topology_updates);
956 #ifdef CONFIG_MEMORY_HOTPLUG
958 * Find the node associated with a hot added memory section for
959 * memory represented in the device tree by the property
960 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
962 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
964 struct drmem_lmb *lmb;
965 unsigned long lmb_size;
966 int nid = NUMA_NO_NODE;
968 lmb_size = drmem_lmb_size();
970 for_each_drmem_lmb(lmb) {
971 /* skip this block if it is reserved or not assigned to
973 if ((lmb->flags & DRCONF_MEM_RESERVED)
974 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
977 if ((scn_addr < lmb->base_addr)
978 || (scn_addr >= (lmb->base_addr + lmb_size)))
981 nid = of_drconf_to_nid_single(lmb);
989 * Find the node associated with a hot added memory section for memory
990 * represented in the device tree as a node (i.e. memory@XXXX) for
993 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
995 struct device_node *memory;
996 int nid = NUMA_NO_NODE;
998 for_each_node_by_type(memory, "memory") {
999 unsigned long start, size;
1001 const __be32 *memcell_buf;
1004 memcell_buf = of_get_property(memory, "reg", &len);
1005 if (!memcell_buf || len <= 0)
1008 /* ranges in cell */
1009 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1012 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1013 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1015 if ((scn_addr < start) || (scn_addr >= (start + size)))
1018 nid = of_node_to_nid_single(memory);
1026 of_node_put(memory);
1032 * Find the node associated with a hot added memory section. Section
1033 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1034 * sections are fully contained within a single MEMBLOCK.
1036 int hot_add_scn_to_nid(unsigned long scn_addr)
1038 struct device_node *memory = NULL;
1042 return first_online_node;
1044 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1046 nid = hot_add_drconf_scn_to_nid(scn_addr);
1047 of_node_put(memory);
1049 nid = hot_add_node_scn_to_nid(scn_addr);
1052 if (nid < 0 || !node_possible(nid))
1053 nid = first_online_node;
1058 static u64 hot_add_drconf_memory_max(void)
1060 struct device_node *memory = NULL;
1061 struct device_node *dn = NULL;
1062 const __be64 *lrdr = NULL;
1064 dn = of_find_node_by_path("/rtas");
1066 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1069 return be64_to_cpup(lrdr);
1072 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1074 of_node_put(memory);
1075 return drmem_lmb_memory_max();
1081 * memory_hotplug_max - return max address of memory that may be added
1083 * This is currently only used on systems that support drconfig memory
1086 u64 memory_hotplug_max(void)
1088 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1090 #endif /* CONFIG_MEMORY_HOTPLUG */
1092 /* Virtual Processor Home Node (VPHN) support */
1093 #ifdef CONFIG_PPC_SPLPAR
1094 struct topology_update_data {
1095 struct topology_update_data *next;
1101 #define TOPOLOGY_DEF_TIMER_SECS 60
1103 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1104 static cpumask_t cpu_associativity_changes_mask;
1105 static int vphn_enabled;
1106 static int prrn_enabled;
1107 static void reset_topology_timer(void);
1108 static int topology_timer_secs = 1;
1109 static int topology_inited;
1112 * Change polling interval for associativity changes.
1114 int timed_topology_update(int nsecs)
1118 topology_timer_secs = nsecs;
1120 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1122 reset_topology_timer();
1129 * Store the current values of the associativity change counters in the
1132 static void setup_cpu_associativity_change_counters(void)
1136 /* The VPHN feature supports a maximum of 8 reference points */
1137 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1139 for_each_possible_cpu(cpu) {
1141 u8 *counts = vphn_cpu_change_counts[cpu];
1142 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1144 for (i = 0; i < distance_ref_points_depth; i++)
1145 counts[i] = hypervisor_counts[i];
1150 * The hypervisor maintains a set of 8 associativity change counters in
1151 * the VPA of each cpu that correspond to the associativity levels in the
1152 * ibm,associativity-reference-points property. When an associativity
1153 * level changes, the corresponding counter is incremented.
1155 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1156 * node associativity levels have changed.
1158 * Returns the number of cpus with unhandled associativity changes.
1160 static int update_cpu_associativity_changes_mask(void)
1163 cpumask_t *changes = &cpu_associativity_changes_mask;
1165 for_each_possible_cpu(cpu) {
1167 u8 *counts = vphn_cpu_change_counts[cpu];
1168 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1170 for (i = 0; i < distance_ref_points_depth; i++) {
1171 if (hypervisor_counts[i] != counts[i]) {
1172 counts[i] = hypervisor_counts[i];
1177 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1178 cpu = cpu_last_thread_sibling(cpu);
1182 return cpumask_weight(changes);
1186 * Retrieve the new associativity information for a virtual processor's
1189 static long vphn_get_associativity(unsigned long cpu,
1190 __be32 *associativity)
1194 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1195 VPHN_FLAG_VCPU, associativity);
1199 printk_once(KERN_INFO
1200 "VPHN is not supported. Disabling polling...\n");
1201 stop_topology_update();
1205 "hcall_vphn() experienced a hardware fault "
1206 "preventing VPHN. Disabling polling...\n");
1207 stop_topology_update();
1210 dbg("VPHN hcall succeeded. Reset polling...\n");
1211 timed_topology_update(0);
1218 int find_and_online_cpu_nid(int cpu)
1220 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1223 /* Use associativity from first thread for all siblings */
1224 if (vphn_get_associativity(cpu, associativity))
1225 return cpu_to_node(cpu);
1227 new_nid = associativity_to_nid(associativity);
1228 if (new_nid < 0 || !node_possible(new_nid))
1229 new_nid = first_online_node;
1231 if (NODE_DATA(new_nid) == NULL) {
1232 #ifdef CONFIG_MEMORY_HOTPLUG
1234 * Need to ensure that NODE_DATA is initialized for a node from
1235 * available memory (see memblock_alloc_try_nid). If unable to
1236 * init the node, then default to nearest node that has memory
1237 * installed. Skip onlining a node if the subsystems are not
1240 if (!topology_inited || try_online_node(new_nid))
1241 new_nid = first_online_node;
1244 * Default to using the nearest node that has memory installed.
1245 * Otherwise, it would be necessary to patch the kernel MM code
1246 * to deal with more memoryless-node error conditions.
1248 new_nid = first_online_node;
1252 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1258 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1259 * characteristics change. This function doesn't perform any locking and is
1260 * only safe to call from stop_machine().
1262 static int update_cpu_topology(void *data)
1264 struct topology_update_data *update;
1270 cpu = smp_processor_id();
1272 for (update = data; update; update = update->next) {
1273 int new_nid = update->new_nid;
1274 if (cpu != update->cpu)
1277 unmap_cpu_from_node(cpu);
1278 map_cpu_to_node(cpu, new_nid);
1279 set_cpu_numa_node(cpu, new_nid);
1280 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1287 static int update_lookup_table(void *data)
1289 struct topology_update_data *update;
1295 * Upon topology update, the numa-cpu lookup table needs to be updated
1296 * for all threads in the core, including offline CPUs, to ensure that
1297 * future hotplug operations respect the cpu-to-node associativity
1300 for (update = data; update; update = update->next) {
1303 nid = update->new_nid;
1304 base = cpu_first_thread_sibling(update->cpu);
1306 for (j = 0; j < threads_per_core; j++) {
1307 update_numa_cpu_lookup_table(base + j, nid);
1315 * Update the node maps and sysfs entries for each cpu whose home node
1316 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1318 * cpus_locked says whether we already hold cpu_hotplug_lock.
1320 int numa_update_cpu_topology(bool cpus_locked)
1322 unsigned int cpu, sibling, changed = 0;
1323 struct topology_update_data *updates, *ud;
1324 cpumask_t updated_cpus;
1326 int weight, new_nid, i = 0;
1328 if (!prrn_enabled && !vphn_enabled && topology_inited)
1331 weight = cpumask_weight(&cpu_associativity_changes_mask);
1335 updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL);
1339 cpumask_clear(&updated_cpus);
1341 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1343 * If siblings aren't flagged for changes, updates list
1344 * will be too short. Skip on this update and set for next
1347 if (!cpumask_subset(cpu_sibling_mask(cpu),
1348 &cpu_associativity_changes_mask)) {
1349 pr_info("Sibling bits not set for associativity "
1350 "change, cpu%d\n", cpu);
1351 cpumask_or(&cpu_associativity_changes_mask,
1352 &cpu_associativity_changes_mask,
1353 cpu_sibling_mask(cpu));
1354 cpu = cpu_last_thread_sibling(cpu);
1358 new_nid = find_and_online_cpu_nid(cpu);
1360 if (new_nid == numa_cpu_lookup_table[cpu]) {
1361 cpumask_andnot(&cpu_associativity_changes_mask,
1362 &cpu_associativity_changes_mask,
1363 cpu_sibling_mask(cpu));
1364 dbg("Assoc chg gives same node %d for cpu%d\n",
1366 cpu = cpu_last_thread_sibling(cpu);
1370 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1372 ud->next = &updates[i];
1374 ud->new_nid = new_nid;
1375 ud->old_nid = numa_cpu_lookup_table[sibling];
1376 cpumask_set_cpu(sibling, &updated_cpus);
1378 cpu = cpu_last_thread_sibling(cpu);
1382 * Prevent processing of 'updates' from overflowing array
1383 * where last entry filled in a 'next' pointer.
1386 updates[i-1].next = NULL;
1388 pr_debug("Topology update for the following CPUs:\n");
1389 if (cpumask_weight(&updated_cpus)) {
1390 for (ud = &updates[0]; ud; ud = ud->next) {
1391 pr_debug("cpu %d moving from node %d "
1393 ud->old_nid, ud->new_nid);
1398 * In cases where we have nothing to update (because the updates list
1399 * is too short or because the new topology is same as the old one),
1400 * skip invoking update_cpu_topology() via stop-machine(). This is
1401 * necessary (and not just a fast-path optimization) since stop-machine
1402 * can end up electing a random CPU to run update_cpu_topology(), and
1403 * thus trick us into setting up incorrect cpu-node mappings (since
1404 * 'updates' is kzalloc()'ed).
1406 * And for the similar reason, we will skip all the following updating.
1408 if (!cpumask_weight(&updated_cpus))
1412 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1415 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1418 * Update the numa-cpu lookup table with the new mappings, even for
1419 * offline CPUs. It is best to perform this update from the stop-
1423 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1424 cpumask_of(raw_smp_processor_id()));
1426 stop_machine(update_lookup_table, &updates[0],
1427 cpumask_of(raw_smp_processor_id()));
1429 for (ud = &updates[0]; ud; ud = ud->next) {
1430 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1431 register_cpu_under_node(ud->cpu, ud->new_nid);
1433 dev = get_cpu_device(ud->cpu);
1435 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1436 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1445 int arch_update_cpu_topology(void)
1447 return numa_update_cpu_topology(true);
1450 static void topology_work_fn(struct work_struct *work)
1452 rebuild_sched_domains();
1454 static DECLARE_WORK(topology_work, topology_work_fn);
1456 static void topology_schedule_update(void)
1458 schedule_work(&topology_work);
1461 static void topology_timer_fn(struct timer_list *unused)
1463 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1464 topology_schedule_update();
1465 else if (vphn_enabled) {
1466 if (update_cpu_associativity_changes_mask() > 0)
1467 topology_schedule_update();
1468 reset_topology_timer();
1471 static struct timer_list topology_timer;
1473 static void reset_topology_timer(void)
1476 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1481 static int dt_update_callback(struct notifier_block *nb,
1482 unsigned long action, void *data)
1484 struct of_reconfig_data *update = data;
1485 int rc = NOTIFY_DONE;
1488 case OF_RECONFIG_UPDATE_PROPERTY:
1489 if (of_node_is_type(update->dn, "cpu") &&
1490 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1492 of_property_read_u32(update->dn, "reg", &core_id);
1493 rc = dlpar_cpu_readd(core_id);
1502 static struct notifier_block dt_update_nb = {
1503 .notifier_call = dt_update_callback,
1509 * Start polling for associativity changes.
1511 int start_topology_update(void)
1515 if (!topology_updates_enabled)
1518 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1519 if (!prrn_enabled) {
1522 rc = of_reconfig_notifier_register(&dt_update_nb);
1526 if (firmware_has_feature(FW_FEATURE_VPHN) &&
1527 lppaca_shared_proc(get_lppaca())) {
1528 if (!vphn_enabled) {
1530 setup_cpu_associativity_change_counters();
1531 timer_setup(&topology_timer, topology_timer_fn,
1533 reset_topology_timer();
1537 pr_info("Starting topology update%s%s\n",
1538 (prrn_enabled ? " prrn_enabled" : ""),
1539 (vphn_enabled ? " vphn_enabled" : ""));
1545 * Disable polling for VPHN associativity changes.
1547 int stop_topology_update(void)
1551 if (!topology_updates_enabled)
1557 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1562 rc = del_timer_sync(&topology_timer);
1565 pr_info("Stopping topology update\n");
1570 int prrn_is_enabled(void)
1572 return prrn_enabled;
1575 void __init shared_proc_topology_init(void)
1577 if (lppaca_shared_proc(get_lppaca())) {
1578 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1580 numa_update_cpu_topology(false);
1584 static int topology_read(struct seq_file *file, void *v)
1586 if (vphn_enabled || prrn_enabled)
1587 seq_puts(file, "on\n");
1589 seq_puts(file, "off\n");
1594 static int topology_open(struct inode *inode, struct file *file)
1596 return single_open(file, topology_read, NULL);
1599 static ssize_t topology_write(struct file *file, const char __user *buf,
1600 size_t count, loff_t *off)
1602 char kbuf[4]; /* "on" or "off" plus null. */
1605 read_len = count < 3 ? count : 3;
1606 if (copy_from_user(kbuf, buf, read_len))
1609 kbuf[read_len] = '\0';
1611 if (!strncmp(kbuf, "on", 2)) {
1612 topology_updates_enabled = true;
1613 start_topology_update();
1614 } else if (!strncmp(kbuf, "off", 3)) {
1615 stop_topology_update();
1616 topology_updates_enabled = false;
1623 static const struct file_operations topology_ops = {
1625 .write = topology_write,
1626 .open = topology_open,
1627 .release = single_release
1630 static int topology_update_init(void)
1632 start_topology_update();
1635 topology_schedule_update();
1637 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1640 topology_inited = 1;
1643 device_initcall(topology_update_init);
1644 #endif /* CONFIG_PPC_SPLPAR */