1 // SPDX-License-Identifier: GPL-2.0-or-later
5 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
7 #define pr_fmt(fmt) "numa: " fmt
9 #include <linux/threads.h>
10 #include <linux/memblock.h>
11 #include <linux/init.h>
13 #include <linux/mmzone.h>
14 #include <linux/export.h>
15 #include <linux/nodemask.h>
16 #include <linux/cpu.h>
17 #include <linux/notifier.h>
19 #include <linux/pfn.h>
20 #include <linux/cpuset.h>
21 #include <linux/node.h>
22 #include <linux/stop_machine.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/uaccess.h>
26 #include <linux/slab.h>
27 #include <asm/cputhreads.h>
28 #include <asm/sparsemem.h>
31 #include <asm/topology.h>
32 #include <asm/firmware.h>
34 #include <asm/hvcall.h>
35 #include <asm/setup.h>
37 #include <asm/drmem.h>
39 static int numa_enabled = 1;
41 static char *cmdline __initdata;
43 static int numa_debug;
44 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
46 int numa_cpu_lookup_table[NR_CPUS];
47 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
48 struct pglist_data *node_data[MAX_NUMNODES];
50 EXPORT_SYMBOL(numa_cpu_lookup_table);
51 EXPORT_SYMBOL(node_to_cpumask_map);
52 EXPORT_SYMBOL(node_data);
54 static int min_common_depth;
55 static int n_mem_addr_cells, n_mem_size_cells;
56 static int form1_affinity;
58 #define MAX_DISTANCE_REF_POINTS 4
59 static int distance_ref_points_depth;
60 static const __be32 *distance_ref_points;
61 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
64 * Allocate node_to_cpumask_map based on number of available nodes
65 * Requires node_possible_map to be valid.
67 * Note: cpumask_of_node() is not valid until after this is done.
69 static void __init setup_node_to_cpumask_map(void)
73 /* setup nr_node_ids if not done yet */
74 if (nr_node_ids == MAX_NUMNODES)
77 /* allocate the map */
79 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
81 /* cpumask_of_node() will now work */
82 dbg("Node to cpumask map for %u nodes\n", nr_node_ids);
85 static int __init fake_numa_create_new_node(unsigned long end_pfn,
88 unsigned long long mem;
90 static unsigned int fake_nid;
91 static unsigned long long curr_boundary;
94 * Modify node id, iff we started creating NUMA nodes
95 * We want to continue from where we left of the last time
100 * In case there are no more arguments to parse, the
101 * node_id should be the same as the last fake node id
102 * (we've handled this above).
107 mem = memparse(p, &p);
111 if (mem < curr_boundary)
116 if ((end_pfn << PAGE_SHIFT) > mem) {
118 * Skip commas and spaces
120 while (*p == ',' || *p == ' ' || *p == '\t')
126 dbg("created new fake_node with id %d\n", fake_nid);
132 static void reset_numa_cpu_lookup_table(void)
136 for_each_possible_cpu(cpu)
137 numa_cpu_lookup_table[cpu] = -1;
140 static void map_cpu_to_node(int cpu, int node)
142 update_numa_cpu_lookup_table(cpu, node);
144 dbg("adding cpu %d to node %d\n", cpu, node);
146 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
147 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
150 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
151 static void unmap_cpu_from_node(unsigned long cpu)
153 int node = numa_cpu_lookup_table[cpu];
155 dbg("removing cpu %lu from node %d\n", cpu, node);
157 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
158 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
160 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
164 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
166 int cpu_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
172 for (i = 0; i < distance_ref_points_depth; i++) {
173 index = be32_to_cpu(distance_ref_points[i]);
174 if (cpu1_assoc[index] == cpu2_assoc[index])
182 /* must hold reference to node during call */
183 static const __be32 *of_get_associativity(struct device_node *dev)
185 return of_get_property(dev, "ibm,associativity", NULL);
188 int __node_distance(int a, int b)
191 int distance = LOCAL_DISTANCE;
194 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
196 for (i = 0; i < distance_ref_points_depth; i++) {
197 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
200 /* Double the distance for each NUMA level */
206 EXPORT_SYMBOL(__node_distance);
208 static void initialize_distance_lookup_table(int nid,
209 const __be32 *associativity)
216 for (i = 0; i < distance_ref_points_depth; i++) {
219 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
220 distance_lookup_table[nid][i] = of_read_number(entry, 1);
225 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
228 static int associativity_to_nid(const __be32 *associativity)
230 int nid = NUMA_NO_NODE;
235 if (of_read_number(associativity, 1) >= min_common_depth)
236 nid = of_read_number(&associativity[min_common_depth], 1);
238 /* POWER4 LPAR uses 0xffff as invalid node */
239 if (nid == 0xffff || nid >= nr_node_ids)
243 of_read_number(associativity, 1) >= distance_ref_points_depth) {
245 * Skip the length field and send start of associativity array
247 initialize_distance_lookup_table(nid, associativity + 1);
254 /* Returns the nid associated with the given device tree node,
255 * or -1 if not found.
257 static int of_node_to_nid_single(struct device_node *device)
259 int nid = NUMA_NO_NODE;
262 tmp = of_get_associativity(device);
264 nid = associativity_to_nid(tmp);
268 /* Walk the device tree upwards, looking for an associativity id */
269 int of_node_to_nid(struct device_node *device)
271 int nid = NUMA_NO_NODE;
275 nid = of_node_to_nid_single(device);
279 device = of_get_next_parent(device);
285 EXPORT_SYMBOL(of_node_to_nid);
287 static int __init find_min_common_depth(void)
290 struct device_node *root;
292 if (firmware_has_feature(FW_FEATURE_OPAL))
293 root = of_find_node_by_path("/ibm,opal");
295 root = of_find_node_by_path("/rtas");
297 root = of_find_node_by_path("/");
300 * This property is a set of 32-bit integers, each representing
301 * an index into the ibm,associativity nodes.
303 * With form 0 affinity the first integer is for an SMP configuration
304 * (should be all 0's) and the second is for a normal NUMA
305 * configuration. We have only one level of NUMA.
307 * With form 1 affinity the first integer is the most significant
308 * NUMA boundary and the following are progressively less significant
309 * boundaries. There can be more than one level of NUMA.
311 distance_ref_points = of_get_property(root,
312 "ibm,associativity-reference-points",
313 &distance_ref_points_depth);
315 if (!distance_ref_points) {
316 dbg("NUMA: ibm,associativity-reference-points not found.\n");
320 distance_ref_points_depth /= sizeof(int);
322 if (firmware_has_feature(FW_FEATURE_OPAL) ||
323 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
324 dbg("Using form 1 affinity\n");
328 if (form1_affinity) {
329 depth = of_read_number(distance_ref_points, 1);
331 if (distance_ref_points_depth < 2) {
332 printk(KERN_WARNING "NUMA: "
333 "short ibm,associativity-reference-points\n");
337 depth = of_read_number(&distance_ref_points[1], 1);
341 * Warn and cap if the hardware supports more than
342 * MAX_DISTANCE_REF_POINTS domains.
344 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
345 printk(KERN_WARNING "NUMA: distance array capped at "
346 "%d entries\n", MAX_DISTANCE_REF_POINTS);
347 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
358 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
360 struct device_node *memory = NULL;
362 memory = of_find_node_by_type(memory, "memory");
364 panic("numa.c: No memory nodes found!");
366 *n_addr_cells = of_n_addr_cells(memory);
367 *n_size_cells = of_n_size_cells(memory);
371 static unsigned long read_n_cells(int n, const __be32 **buf)
373 unsigned long result = 0;
376 result = (result << 32) | of_read_number(*buf, 1);
382 struct assoc_arrays {
385 const __be32 *arrays;
389 * Retrieve and validate the list of associativity arrays for drconf
390 * memory from the ibm,associativity-lookup-arrays property of the
393 * The layout of the ibm,associativity-lookup-arrays property is a number N
394 * indicating the number of associativity arrays, followed by a number M
395 * indicating the size of each associativity array, followed by a list
396 * of N associativity arrays.
398 static int of_get_assoc_arrays(struct assoc_arrays *aa)
400 struct device_node *memory;
404 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
408 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
409 if (!prop || len < 2 * sizeof(unsigned int)) {
414 aa->n_arrays = of_read_number(prop++, 1);
415 aa->array_sz = of_read_number(prop++, 1);
419 /* Now that we know the number of arrays and size of each array,
420 * revalidate the size of the property read in.
422 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
430 * This is like of_node_to_nid_single() for memory represented in the
431 * ibm,dynamic-reconfiguration-memory node.
433 int of_drconf_to_nid_single(struct drmem_lmb *lmb)
435 struct assoc_arrays aa = { .arrays = NULL };
436 int default_nid = NUMA_NO_NODE;
437 int nid = default_nid;
440 if ((min_common_depth < 0) || !numa_enabled)
443 rc = of_get_assoc_arrays(&aa);
447 if (min_common_depth <= aa.array_sz &&
448 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
449 index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
450 nid = of_read_number(&aa.arrays[index], 1);
452 if (nid == 0xffff || nid >= nr_node_ids)
456 index = lmb->aa_index * aa.array_sz;
457 initialize_distance_lookup_table(nid,
465 #ifdef CONFIG_PPC_SPLPAR
466 static int vphn_get_nid(long lcpu)
468 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
472 * On a shared lpar, device tree will not have node associativity.
473 * At this time lppaca, or its __old_status field may not be
474 * updated. Hence kernel cannot detect if its on a shared lpar. So
475 * request an explicit associativity irrespective of whether the
476 * lpar is shared or dedicated. Use the device tree property as a
477 * fallback. cpu_to_phys_id is only valid between
478 * smp_setup_cpu_maps() and smp_setup_pacas().
480 if (firmware_has_feature(FW_FEATURE_VPHN)) {
482 hwid = cpu_to_phys_id[lcpu];
484 hwid = get_hard_smp_processor_id(lcpu);
486 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
488 return associativity_to_nid(associativity);
494 static int vphn_get_nid(long unused)
498 #endif /* CONFIG_PPC_SPLPAR */
501 * Figure out to which domain a cpu belongs and stick it there.
502 * Return the id of the domain used.
504 static int numa_setup_cpu(unsigned long lcpu)
506 struct device_node *cpu;
507 int fcpu = cpu_first_thread_sibling(lcpu);
508 int nid = NUMA_NO_NODE;
510 if (!cpu_present(lcpu)) {
511 set_cpu_numa_node(lcpu, first_online_node);
512 return first_online_node;
516 * If a valid cpu-to-node mapping is already available, use it
517 * directly instead of querying the firmware, since it represents
518 * the most recent mapping notified to us by the platform (eg: VPHN).
519 * Since cpu_to_node binding remains the same for all threads in the
520 * core. If a valid cpu-to-node mapping is already available, for
521 * the first thread in the core, use it.
523 nid = numa_cpu_lookup_table[fcpu];
525 map_cpu_to_node(lcpu, nid);
529 nid = vphn_get_nid(lcpu);
530 if (nid != NUMA_NO_NODE)
533 cpu = of_get_cpu_node(lcpu, NULL);
537 if (cpu_present(lcpu))
543 nid = of_node_to_nid_single(cpu);
547 if (nid < 0 || !node_possible(nid))
548 nid = first_online_node;
551 * Update for the first thread of the core. All threads of a core
552 * have to be part of the same node. This not only avoids querying
553 * for every other thread in the core, but always avoids a case
554 * where virtual node associativity change causes subsequent threads
555 * of a core to be associated with different nid. However if first
556 * thread is already online, expect it to have a valid mapping.
559 WARN_ON(cpu_online(fcpu));
560 map_cpu_to_node(fcpu, nid);
563 map_cpu_to_node(lcpu, nid);
568 static void verify_cpu_node_mapping(int cpu, int node)
570 int base, sibling, i;
572 /* Verify that all the threads in the core belong to the same node */
573 base = cpu_first_thread_sibling(cpu);
575 for (i = 0; i < threads_per_core; i++) {
578 if (sibling == cpu || cpu_is_offline(sibling))
581 if (cpu_to_node(sibling) != node) {
582 WARN(1, "CPU thread siblings %d and %d don't belong"
583 " to the same node!\n", cpu, sibling);
589 /* Must run before sched domains notifier. */
590 static int ppc_numa_cpu_prepare(unsigned int cpu)
594 nid = numa_setup_cpu(cpu);
595 verify_cpu_node_mapping(cpu, nid);
599 static int ppc_numa_cpu_dead(unsigned int cpu)
601 #ifdef CONFIG_HOTPLUG_CPU
602 unmap_cpu_from_node(cpu);
608 * Check and possibly modify a memory region to enforce the memory limit.
610 * Returns the size the region should have to enforce the memory limit.
611 * This will either be the original value of size, a truncated value,
612 * or zero. If the returned value of size is 0 the region should be
613 * discarded as it lies wholly above the memory limit.
615 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
619 * We use memblock_end_of_DRAM() in here instead of memory_limit because
620 * we've already adjusted it for the limit and it takes care of
621 * having memory holes below the limit. Also, in the case of
622 * iommu_is_off, memory_limit is not set but is implicitly enforced.
625 if (start + size <= memblock_end_of_DRAM())
628 if (start >= memblock_end_of_DRAM())
631 return memblock_end_of_DRAM() - start;
635 * Reads the counter for a given entry in
636 * linux,drconf-usable-memory property
638 static inline int __init read_usm_ranges(const __be32 **usm)
641 * For each lmb in ibm,dynamic-memory a corresponding
642 * entry in linux,drconf-usable-memory property contains
643 * a counter followed by that many (base, size) duple.
644 * read the counter from linux,drconf-usable-memory
646 return read_n_cells(n_mem_size_cells, usm);
650 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
651 * node. This assumes n_mem_{addr,size}_cells have been set.
653 static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
657 unsigned int ranges, is_kexec_kdump = 0;
658 unsigned long base, size, sz;
662 * Skip this block if the reserved bit is set in flags (0x80)
663 * or if the block is not assigned to this partition (0x8)
665 if ((lmb->flags & DRCONF_MEM_RESERVED)
666 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
672 base = lmb->base_addr;
673 size = drmem_lmb_size();
676 if (is_kexec_kdump) {
677 ranges = read_usm_ranges(usm);
678 if (!ranges) /* there are no (base, size) duple */
683 if (is_kexec_kdump) {
684 base = read_n_cells(n_mem_addr_cells, usm);
685 size = read_n_cells(n_mem_size_cells, usm);
688 nid = of_drconf_to_nid_single(lmb);
689 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
691 node_set_online(nid);
692 sz = numa_enforce_memory_limit(base, size);
694 memblock_set_node(base, sz, &memblock.memory, nid);
700 static int __init parse_numa_properties(void)
702 struct device_node *memory;
706 if (numa_enabled == 0) {
707 printk(KERN_WARNING "NUMA disabled by user\n");
711 min_common_depth = find_min_common_depth();
713 if (min_common_depth < 0) {
715 * if we fail to parse min_common_depth from device tree
716 * mark the numa disabled, boot with numa disabled.
718 numa_enabled = false;
719 return min_common_depth;
722 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
725 * Even though we connect cpus to numa domains later in SMP
726 * init, we need to know the node ids now. This is because
727 * each node to be onlined must have NODE_DATA etc backing it.
729 for_each_present_cpu(i) {
730 struct device_node *cpu;
731 int nid = vphn_get_nid(i);
734 * Don't fall back to default_nid yet -- we will plug
735 * cpus into nodes once the memory scan has discovered
738 if (nid == NUMA_NO_NODE) {
739 cpu = of_get_cpu_node(i, NULL);
741 nid = of_node_to_nid_single(cpu);
746 node_set_online(nid);
749 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
751 for_each_node_by_type(memory, "memory") {
756 const __be32 *memcell_buf;
759 memcell_buf = of_get_property(memory,
760 "linux,usable-memory", &len);
761 if (!memcell_buf || len <= 0)
762 memcell_buf = of_get_property(memory, "reg", &len);
763 if (!memcell_buf || len <= 0)
767 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
769 /* these are order-sensitive, and modify the buffer pointer */
770 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
771 size = read_n_cells(n_mem_size_cells, &memcell_buf);
774 * Assumption: either all memory nodes or none will
775 * have associativity properties. If none, then
776 * everything goes to default_nid.
778 nid = of_node_to_nid_single(memory);
782 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
783 node_set_online(nid);
785 size = numa_enforce_memory_limit(start, size);
787 memblock_set_node(start, size, &memblock.memory, nid);
794 * Now do the same thing for each MEMBLOCK listed in the
795 * ibm,dynamic-memory property in the
796 * ibm,dynamic-reconfiguration-memory node.
798 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
800 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
807 static void __init setup_nonnuma(void)
809 unsigned long top_of_ram = memblock_end_of_DRAM();
810 unsigned long total_ram = memblock_phys_mem_size();
811 unsigned long start_pfn, end_pfn;
812 unsigned int nid = 0;
815 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
816 top_of_ram, total_ram);
817 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
818 (top_of_ram - total_ram) >> 20);
820 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
821 fake_numa_create_new_node(end_pfn, &nid);
822 memblock_set_node(PFN_PHYS(start_pfn),
823 PFN_PHYS(end_pfn - start_pfn),
824 &memblock.memory, nid);
825 node_set_online(nid);
829 void __init dump_numa_cpu_topology(void)
832 unsigned int cpu, count;
837 for_each_online_node(node) {
838 pr_info("Node %d CPUs:", node);
842 * If we used a CPU iterator here we would miss printing
843 * the holes in the cpumap.
845 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
846 if (cpumask_test_cpu(cpu,
847 node_to_cpumask_map[node])) {
853 pr_cont("-%u", cpu - 1);
859 pr_cont("-%u", nr_cpu_ids - 1);
864 /* Initialize NODE_DATA for a node on the local memory */
865 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
867 u64 spanned_pages = end_pfn - start_pfn;
868 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
873 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
875 panic("Cannot allocate %zu bytes for node %d data\n",
880 /* report and initialize */
881 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
882 nd_pa, nd_pa + nd_size - 1);
883 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
885 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
888 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
889 NODE_DATA(nid)->node_id = nid;
890 NODE_DATA(nid)->node_start_pfn = start_pfn;
891 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
894 static void __init find_possible_nodes(void)
896 struct device_node *rtas;
897 const __be32 *domains;
898 int prop_length, max_nodes;
904 rtas = of_find_node_by_path("/rtas");
909 * ibm,current-associativity-domains is a fairly recent property. If
910 * it doesn't exist, then fallback on ibm,max-associativity-domains.
911 * Current denotes what the platform can support compared to max
912 * which denotes what the Hypervisor can support.
914 domains = of_get_property(rtas, "ibm,current-associativity-domains",
917 domains = of_get_property(rtas, "ibm,max-associativity-domains",
923 max_nodes = of_read_number(&domains[min_common_depth], 1);
924 for (i = 0; i < max_nodes; i++) {
925 if (!node_possible(i))
926 node_set(i, node_possible_map);
929 prop_length /= sizeof(int);
930 if (prop_length > min_common_depth + 2)
931 coregroup_enabled = 1;
937 void __init mem_topology_setup(void)
942 * Linux/mm assumes node 0 to be online at boot. However this is not
943 * true on PowerPC, where node 0 is similar to any other node, it
944 * could be cpuless, memoryless node. So force node 0 to be offline
945 * for now. This will prevent cpuless, memoryless node 0 showing up
946 * unnecessarily as online. If a node has cpus or memory that need
947 * to be online, then node will anyway be marked online.
951 if (parse_numa_properties())
955 * Modify the set of possible NUMA nodes to reflect information
956 * available about the set of online nodes, and the set of nodes
957 * that we expect to make use of for this platform's affinity
960 nodes_and(node_possible_map, node_possible_map, node_online_map);
962 find_possible_nodes();
964 setup_node_to_cpumask_map();
966 reset_numa_cpu_lookup_table();
968 for_each_possible_cpu(cpu) {
970 * Powerpc with CONFIG_NUMA always used to have a node 0,
971 * even if it was memoryless or cpuless. For all cpus that
972 * are possible but not present, cpu_to_node() would point
973 * to node 0. To remove a cpuless, memoryless dummy node,
974 * powerpc need to make sure all possible but not present
975 * cpu_to_node are set to a proper node.
981 void __init initmem_init(void)
985 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
986 max_pfn = max_low_pfn;
990 for_each_online_node(nid) {
991 unsigned long start_pfn, end_pfn;
993 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
994 setup_node_data(nid, start_pfn, end_pfn);
1000 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1001 * even before we online them, so that we can use cpu_to_{node,mem}
1002 * early in boot, cf. smp_prepare_cpus().
1003 * _nocalls() + manual invocation is used because cpuhp is not yet
1004 * initialized for the boot CPU.
1006 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1007 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
1010 static int __init early_numa(char *p)
1015 if (strstr(p, "off"))
1018 if (strstr(p, "debug"))
1021 p = strstr(p, "fake=");
1023 cmdline = p + strlen("fake=");
1027 early_param("numa", early_numa);
1029 #ifdef CONFIG_MEMORY_HOTPLUG
1031 * Find the node associated with a hot added memory section for
1032 * memory represented in the device tree by the property
1033 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1035 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
1037 struct drmem_lmb *lmb;
1038 unsigned long lmb_size;
1039 int nid = NUMA_NO_NODE;
1041 lmb_size = drmem_lmb_size();
1043 for_each_drmem_lmb(lmb) {
1044 /* skip this block if it is reserved or not assigned to
1046 if ((lmb->flags & DRCONF_MEM_RESERVED)
1047 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1050 if ((scn_addr < lmb->base_addr)
1051 || (scn_addr >= (lmb->base_addr + lmb_size)))
1054 nid = of_drconf_to_nid_single(lmb);
1062 * Find the node associated with a hot added memory section for memory
1063 * represented in the device tree as a node (i.e. memory@XXXX) for
1066 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1068 struct device_node *memory;
1069 int nid = NUMA_NO_NODE;
1071 for_each_node_by_type(memory, "memory") {
1072 unsigned long start, size;
1074 const __be32 *memcell_buf;
1077 memcell_buf = of_get_property(memory, "reg", &len);
1078 if (!memcell_buf || len <= 0)
1081 /* ranges in cell */
1082 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1085 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1086 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1088 if ((scn_addr < start) || (scn_addr >= (start + size)))
1091 nid = of_node_to_nid_single(memory);
1099 of_node_put(memory);
1105 * Find the node associated with a hot added memory section. Section
1106 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1107 * sections are fully contained within a single MEMBLOCK.
1109 int hot_add_scn_to_nid(unsigned long scn_addr)
1111 struct device_node *memory = NULL;
1115 return first_online_node;
1117 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1119 nid = hot_add_drconf_scn_to_nid(scn_addr);
1120 of_node_put(memory);
1122 nid = hot_add_node_scn_to_nid(scn_addr);
1125 if (nid < 0 || !node_possible(nid))
1126 nid = first_online_node;
1131 static u64 hot_add_drconf_memory_max(void)
1133 struct device_node *memory = NULL;
1134 struct device_node *dn = NULL;
1135 const __be64 *lrdr = NULL;
1137 dn = of_find_node_by_path("/rtas");
1139 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1142 return be64_to_cpup(lrdr);
1145 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1147 of_node_put(memory);
1148 return drmem_lmb_memory_max();
1154 * memory_hotplug_max - return max address of memory that may be added
1156 * This is currently only used on systems that support drconfig memory
1159 u64 memory_hotplug_max(void)
1161 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1163 #endif /* CONFIG_MEMORY_HOTPLUG */
1165 /* Virtual Processor Home Node (VPHN) support */
1166 #ifdef CONFIG_PPC_SPLPAR
1167 static int topology_inited;
1170 * Retrieve the new associativity information for a virtual processor's
1173 static long vphn_get_associativity(unsigned long cpu,
1174 __be32 *associativity)
1178 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1179 VPHN_FLAG_VCPU, associativity);
1183 dbg("VPHN hcall succeeded. Reset polling...\n");
1187 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
1190 pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
1191 "preventing VPHN. Disabling polling...\n");
1194 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1195 "Disabling polling...\n");
1198 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1206 int find_and_online_cpu_nid(int cpu)
1208 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1211 /* Use associativity from first thread for all siblings */
1212 if (vphn_get_associativity(cpu, associativity))
1213 return cpu_to_node(cpu);
1215 new_nid = associativity_to_nid(associativity);
1216 if (new_nid < 0 || !node_possible(new_nid))
1217 new_nid = first_online_node;
1219 if (NODE_DATA(new_nid) == NULL) {
1220 #ifdef CONFIG_MEMORY_HOTPLUG
1222 * Need to ensure that NODE_DATA is initialized for a node from
1223 * available memory (see memblock_alloc_try_nid). If unable to
1224 * init the node, then default to nearest node that has memory
1225 * installed. Skip onlining a node if the subsystems are not
1228 if (!topology_inited || try_online_node(new_nid))
1229 new_nid = first_online_node;
1232 * Default to using the nearest node that has memory installed.
1233 * Otherwise, it would be necessary to patch the kernel MM code
1234 * to deal with more memoryless-node error conditions.
1236 new_nid = first_online_node;
1240 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1245 int cpu_to_coregroup_id(int cpu)
1247 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1250 if (cpu < 0 || cpu > nr_cpu_ids)
1253 if (!coregroup_enabled)
1256 if (!firmware_has_feature(FW_FEATURE_VPHN))
1259 if (vphn_get_associativity(cpu, associativity))
1262 index = of_read_number(associativity, 1);
1263 if (index > min_common_depth + 1)
1264 return of_read_number(&associativity[index - 1], 1);
1267 return cpu_to_core_id(cpu);
1270 static int topology_update_init(void)
1272 topology_inited = 1;
1275 device_initcall(topology_update_init);
1276 #endif /* CONFIG_PPC_SPLPAR */