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);
224 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
227 static int associativity_to_nid(const __be32 *associativity)
229 int nid = NUMA_NO_NODE;
234 if (of_read_number(associativity, 1) >= min_common_depth)
235 nid = of_read_number(&associativity[min_common_depth], 1);
237 /* POWER4 LPAR uses 0xffff as invalid node */
238 if (nid == 0xffff || nid >= MAX_NUMNODES)
242 of_read_number(associativity, 1) >= distance_ref_points_depth) {
244 * Skip the length field and send start of associativity array
246 initialize_distance_lookup_table(nid, associativity + 1);
253 /* Returns the nid associated with the given device tree node,
254 * or -1 if not found.
256 static int of_node_to_nid_single(struct device_node *device)
258 int nid = NUMA_NO_NODE;
261 tmp = of_get_associativity(device);
263 nid = associativity_to_nid(tmp);
267 /* Walk the device tree upwards, looking for an associativity id */
268 int of_node_to_nid(struct device_node *device)
270 int nid = NUMA_NO_NODE;
274 nid = of_node_to_nid_single(device);
278 device = of_get_next_parent(device);
284 EXPORT_SYMBOL(of_node_to_nid);
286 static int __init find_min_common_depth(void)
289 struct device_node *root;
291 if (firmware_has_feature(FW_FEATURE_OPAL))
292 root = of_find_node_by_path("/ibm,opal");
294 root = of_find_node_by_path("/rtas");
296 root = of_find_node_by_path("/");
299 * This property is a set of 32-bit integers, each representing
300 * an index into the ibm,associativity nodes.
302 * With form 0 affinity the first integer is for an SMP configuration
303 * (should be all 0's) and the second is for a normal NUMA
304 * configuration. We have only one level of NUMA.
306 * With form 1 affinity the first integer is the most significant
307 * NUMA boundary and the following are progressively less significant
308 * boundaries. There can be more than one level of NUMA.
310 distance_ref_points = of_get_property(root,
311 "ibm,associativity-reference-points",
312 &distance_ref_points_depth);
314 if (!distance_ref_points) {
315 dbg("NUMA: ibm,associativity-reference-points not found.\n");
319 distance_ref_points_depth /= sizeof(int);
321 if (firmware_has_feature(FW_FEATURE_OPAL) ||
322 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
323 dbg("Using form 1 affinity\n");
327 if (form1_affinity) {
328 depth = of_read_number(distance_ref_points, 1);
330 if (distance_ref_points_depth < 2) {
331 printk(KERN_WARNING "NUMA: "
332 "short ibm,associativity-reference-points\n");
336 depth = of_read_number(&distance_ref_points[1], 1);
340 * Warn and cap if the hardware supports more than
341 * MAX_DISTANCE_REF_POINTS domains.
343 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
344 printk(KERN_WARNING "NUMA: distance array capped at "
345 "%d entries\n", MAX_DISTANCE_REF_POINTS);
346 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
357 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
359 struct device_node *memory = NULL;
361 memory = of_find_node_by_type(memory, "memory");
363 panic("numa.c: No memory nodes found!");
365 *n_addr_cells = of_n_addr_cells(memory);
366 *n_size_cells = of_n_size_cells(memory);
370 static unsigned long read_n_cells(int n, const __be32 **buf)
372 unsigned long result = 0;
375 result = (result << 32) | of_read_number(*buf, 1);
381 struct assoc_arrays {
384 const __be32 *arrays;
388 * Retrieve and validate the list of associativity arrays for drconf
389 * memory from the ibm,associativity-lookup-arrays property of the
392 * The layout of the ibm,associativity-lookup-arrays property is a number N
393 * indicating the number of associativity arrays, followed by a number M
394 * indicating the size of each associativity array, followed by a list
395 * of N associativity arrays.
397 static int of_get_assoc_arrays(struct assoc_arrays *aa)
399 struct device_node *memory;
403 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
407 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
408 if (!prop || len < 2 * sizeof(unsigned int)) {
413 aa->n_arrays = of_read_number(prop++, 1);
414 aa->array_sz = of_read_number(prop++, 1);
418 /* Now that we know the number of arrays and size of each array,
419 * revalidate the size of the property read in.
421 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
429 * This is like of_node_to_nid_single() for memory represented in the
430 * ibm,dynamic-reconfiguration-memory node.
432 static int of_drconf_to_nid_single(struct drmem_lmb *lmb)
434 struct assoc_arrays aa = { .arrays = NULL };
435 int default_nid = NUMA_NO_NODE;
436 int nid = default_nid;
439 if ((min_common_depth < 0) || !numa_enabled)
442 rc = of_get_assoc_arrays(&aa);
446 if (min_common_depth <= aa.array_sz &&
447 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
448 index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
449 nid = of_read_number(&aa.arrays[index], 1);
451 if (nid == 0xffff || nid >= MAX_NUMNODES)
455 index = lmb->aa_index * aa.array_sz;
456 initialize_distance_lookup_table(nid,
465 * Figure out to which domain a cpu belongs and stick it there.
466 * Return the id of the domain used.
468 static int numa_setup_cpu(unsigned long lcpu)
470 int nid = NUMA_NO_NODE;
471 struct device_node *cpu;
474 * If a valid cpu-to-node mapping is already available, use it
475 * directly instead of querying the firmware, since it represents
476 * the most recent mapping notified to us by the platform (eg: VPHN).
478 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
479 map_cpu_to_node(lcpu, nid);
483 cpu = of_get_cpu_node(lcpu, NULL);
487 if (cpu_present(lcpu))
493 nid = of_node_to_nid_single(cpu);
496 if (nid < 0 || !node_possible(nid))
497 nid = first_online_node;
499 map_cpu_to_node(lcpu, nid);
505 static void verify_cpu_node_mapping(int cpu, int node)
507 int base, sibling, i;
509 /* Verify that all the threads in the core belong to the same node */
510 base = cpu_first_thread_sibling(cpu);
512 for (i = 0; i < threads_per_core; i++) {
515 if (sibling == cpu || cpu_is_offline(sibling))
518 if (cpu_to_node(sibling) != node) {
519 WARN(1, "CPU thread siblings %d and %d don't belong"
520 " to the same node!\n", cpu, sibling);
526 /* Must run before sched domains notifier. */
527 static int ppc_numa_cpu_prepare(unsigned int cpu)
531 nid = numa_setup_cpu(cpu);
532 verify_cpu_node_mapping(cpu, nid);
536 static int ppc_numa_cpu_dead(unsigned int cpu)
538 #ifdef CONFIG_HOTPLUG_CPU
539 unmap_cpu_from_node(cpu);
545 * Check and possibly modify a memory region to enforce the memory limit.
547 * Returns the size the region should have to enforce the memory limit.
548 * This will either be the original value of size, a truncated value,
549 * or zero. If the returned value of size is 0 the region should be
550 * discarded as it lies wholly above the memory limit.
552 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
556 * We use memblock_end_of_DRAM() in here instead of memory_limit because
557 * we've already adjusted it for the limit and it takes care of
558 * having memory holes below the limit. Also, in the case of
559 * iommu_is_off, memory_limit is not set but is implicitly enforced.
562 if (start + size <= memblock_end_of_DRAM())
565 if (start >= memblock_end_of_DRAM())
568 return memblock_end_of_DRAM() - start;
572 * Reads the counter for a given entry in
573 * linux,drconf-usable-memory property
575 static inline int __init read_usm_ranges(const __be32 **usm)
578 * For each lmb in ibm,dynamic-memory a corresponding
579 * entry in linux,drconf-usable-memory property contains
580 * a counter followed by that many (base, size) duple.
581 * read the counter from linux,drconf-usable-memory
583 return read_n_cells(n_mem_size_cells, usm);
587 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
588 * node. This assumes n_mem_{addr,size}_cells have been set.
590 static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
593 unsigned int ranges, is_kexec_kdump = 0;
594 unsigned long base, size, sz;
598 * Skip this block if the reserved bit is set in flags (0x80)
599 * or if the block is not assigned to this partition (0x8)
601 if ((lmb->flags & DRCONF_MEM_RESERVED)
602 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
608 base = lmb->base_addr;
609 size = drmem_lmb_size();
612 if (is_kexec_kdump) {
613 ranges = read_usm_ranges(usm);
614 if (!ranges) /* there are no (base, size) duple */
619 if (is_kexec_kdump) {
620 base = read_n_cells(n_mem_addr_cells, usm);
621 size = read_n_cells(n_mem_size_cells, usm);
624 nid = of_drconf_to_nid_single(lmb);
625 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
627 node_set_online(nid);
628 sz = numa_enforce_memory_limit(base, size);
630 memblock_set_node(base, sz, &memblock.memory, nid);
634 static int __init parse_numa_properties(void)
636 struct device_node *memory;
640 if (numa_enabled == 0) {
641 printk(KERN_WARNING "NUMA disabled by user\n");
645 min_common_depth = find_min_common_depth();
647 if (min_common_depth < 0) {
649 * if we fail to parse min_common_depth from device tree
650 * mark the numa disabled, boot with numa disabled.
652 numa_enabled = false;
653 return min_common_depth;
656 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
659 * Even though we connect cpus to numa domains later in SMP
660 * init, we need to know the node ids now. This is because
661 * each node to be onlined must have NODE_DATA etc backing it.
663 for_each_present_cpu(i) {
664 struct device_node *cpu;
667 cpu = of_get_cpu_node(i, NULL);
669 nid = of_node_to_nid_single(cpu);
673 * Don't fall back to default_nid yet -- we will plug
674 * cpus into nodes once the memory scan has discovered
679 node_set_online(nid);
682 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
684 for_each_node_by_type(memory, "memory") {
689 const __be32 *memcell_buf;
692 memcell_buf = of_get_property(memory,
693 "linux,usable-memory", &len);
694 if (!memcell_buf || len <= 0)
695 memcell_buf = of_get_property(memory, "reg", &len);
696 if (!memcell_buf || len <= 0)
700 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
702 /* these are order-sensitive, and modify the buffer pointer */
703 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
704 size = read_n_cells(n_mem_size_cells, &memcell_buf);
707 * Assumption: either all memory nodes or none will
708 * have associativity properties. If none, then
709 * everything goes to default_nid.
711 nid = of_node_to_nid_single(memory);
715 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
716 node_set_online(nid);
718 size = numa_enforce_memory_limit(start, size);
720 memblock_set_node(start, size, &memblock.memory, nid);
727 * Now do the same thing for each MEMBLOCK listed in the
728 * ibm,dynamic-memory property in the
729 * ibm,dynamic-reconfiguration-memory node.
731 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
733 walk_drmem_lmbs(memory, numa_setup_drmem_lmb);
740 static void __init setup_nonnuma(void)
742 unsigned long top_of_ram = memblock_end_of_DRAM();
743 unsigned long total_ram = memblock_phys_mem_size();
744 unsigned long start_pfn, end_pfn;
745 unsigned int nid = 0;
746 struct memblock_region *reg;
748 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
749 top_of_ram, total_ram);
750 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
751 (top_of_ram - total_ram) >> 20);
753 for_each_memblock(memory, reg) {
754 start_pfn = memblock_region_memory_base_pfn(reg);
755 end_pfn = memblock_region_memory_end_pfn(reg);
757 fake_numa_create_new_node(end_pfn, &nid);
758 memblock_set_node(PFN_PHYS(start_pfn),
759 PFN_PHYS(end_pfn - start_pfn),
760 &memblock.memory, nid);
761 node_set_online(nid);
765 void __init dump_numa_cpu_topology(void)
768 unsigned int cpu, count;
773 for_each_online_node(node) {
774 pr_info("Node %d CPUs:", node);
778 * If we used a CPU iterator here we would miss printing
779 * the holes in the cpumap.
781 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
782 if (cpumask_test_cpu(cpu,
783 node_to_cpumask_map[node])) {
789 pr_cont("-%u", cpu - 1);
795 pr_cont("-%u", nr_cpu_ids - 1);
800 /* Initialize NODE_DATA for a node on the local memory */
801 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
803 u64 spanned_pages = end_pfn - start_pfn;
804 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
809 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
811 panic("Cannot allocate %zu bytes for node %d data\n",
816 /* report and initialize */
817 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
818 nd_pa, nd_pa + nd_size - 1);
819 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
821 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
824 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
825 NODE_DATA(nid)->node_id = nid;
826 NODE_DATA(nid)->node_start_pfn = start_pfn;
827 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
830 static void __init find_possible_nodes(void)
832 struct device_node *rtas;
838 rtas = of_find_node_by_path("/rtas");
842 if (of_property_read_u32_index(rtas,
843 "ibm,max-associativity-domains",
844 min_common_depth, &numnodes))
847 for (i = 0; i < numnodes; i++) {
848 if (!node_possible(i))
849 node_set(i, node_possible_map);
856 void __init mem_topology_setup(void)
860 if (parse_numa_properties())
864 * Modify the set of possible NUMA nodes to reflect information
865 * available about the set of online nodes, and the set of nodes
866 * that we expect to make use of for this platform's affinity
869 nodes_and(node_possible_map, node_possible_map, node_online_map);
871 find_possible_nodes();
873 setup_node_to_cpumask_map();
875 reset_numa_cpu_lookup_table();
877 for_each_present_cpu(cpu)
881 void __init initmem_init(void)
885 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
886 max_pfn = max_low_pfn;
890 for_each_online_node(nid) {
891 unsigned long start_pfn, end_pfn;
893 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
894 setup_node_data(nid, start_pfn, end_pfn);
895 sparse_memory_present_with_active_regions(nid);
901 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
902 * even before we online them, so that we can use cpu_to_{node,mem}
903 * early in boot, cf. smp_prepare_cpus().
904 * _nocalls() + manual invocation is used because cpuhp is not yet
905 * initialized for the boot CPU.
907 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
908 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
911 static int __init early_numa(char *p)
916 if (strstr(p, "off"))
919 if (strstr(p, "debug"))
922 p = strstr(p, "fake=");
924 cmdline = p + strlen("fake=");
928 early_param("numa", early_numa);
931 * The platform can inform us through one of several mechanisms
932 * (post-migration device tree updates, PRRN or VPHN) that the NUMA
933 * assignment of a resource has changed. This controls whether we act
934 * on that. Disabled by default.
936 static bool topology_updates_enabled;
938 static int __init early_topology_updates(char *p)
943 if (!strcmp(p, "on")) {
944 pr_warn("Caution: enabling topology updates\n");
945 topology_updates_enabled = true;
950 early_param("topology_updates", early_topology_updates);
952 #ifdef CONFIG_MEMORY_HOTPLUG
954 * Find the node associated with a hot added memory section for
955 * memory represented in the device tree by the property
956 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
958 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
960 struct drmem_lmb *lmb;
961 unsigned long lmb_size;
962 int nid = NUMA_NO_NODE;
964 lmb_size = drmem_lmb_size();
966 for_each_drmem_lmb(lmb) {
967 /* skip this block if it is reserved or not assigned to
969 if ((lmb->flags & DRCONF_MEM_RESERVED)
970 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
973 if ((scn_addr < lmb->base_addr)
974 || (scn_addr >= (lmb->base_addr + lmb_size)))
977 nid = of_drconf_to_nid_single(lmb);
985 * Find the node associated with a hot added memory section for memory
986 * represented in the device tree as a node (i.e. memory@XXXX) for
989 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
991 struct device_node *memory;
992 int nid = NUMA_NO_NODE;
994 for_each_node_by_type(memory, "memory") {
995 unsigned long start, size;
997 const __be32 *memcell_buf;
1000 memcell_buf = of_get_property(memory, "reg", &len);
1001 if (!memcell_buf || len <= 0)
1004 /* ranges in cell */
1005 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1008 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1009 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1011 if ((scn_addr < start) || (scn_addr >= (start + size)))
1014 nid = of_node_to_nid_single(memory);
1022 of_node_put(memory);
1028 * Find the node associated with a hot added memory section. Section
1029 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1030 * sections are fully contained within a single MEMBLOCK.
1032 int hot_add_scn_to_nid(unsigned long scn_addr)
1034 struct device_node *memory = NULL;
1038 return first_online_node;
1040 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1042 nid = hot_add_drconf_scn_to_nid(scn_addr);
1043 of_node_put(memory);
1045 nid = hot_add_node_scn_to_nid(scn_addr);
1048 if (nid < 0 || !node_possible(nid))
1049 nid = first_online_node;
1054 static u64 hot_add_drconf_memory_max(void)
1056 struct device_node *memory = NULL;
1057 struct device_node *dn = NULL;
1058 const __be64 *lrdr = NULL;
1060 dn = of_find_node_by_path("/rtas");
1062 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1065 return be64_to_cpup(lrdr);
1068 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1070 of_node_put(memory);
1071 return drmem_lmb_memory_max();
1077 * memory_hotplug_max - return max address of memory that may be added
1079 * This is currently only used on systems that support drconfig memory
1082 u64 memory_hotplug_max(void)
1084 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1086 #endif /* CONFIG_MEMORY_HOTPLUG */
1088 /* Virtual Processor Home Node (VPHN) support */
1089 #ifdef CONFIG_PPC_SPLPAR
1090 struct topology_update_data {
1091 struct topology_update_data *next;
1097 #define TOPOLOGY_DEF_TIMER_SECS 60
1099 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1100 static cpumask_t cpu_associativity_changes_mask;
1101 static int vphn_enabled;
1102 static int prrn_enabled;
1103 static void reset_topology_timer(void);
1104 static int topology_timer_secs = 1;
1105 static int topology_inited;
1108 * Change polling interval for associativity changes.
1110 int timed_topology_update(int nsecs)
1114 topology_timer_secs = nsecs;
1116 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1118 reset_topology_timer();
1125 * Store the current values of the associativity change counters in the
1128 static void setup_cpu_associativity_change_counters(void)
1132 /* The VPHN feature supports a maximum of 8 reference points */
1133 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
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 counts[i] = hypervisor_counts[i];
1146 * The hypervisor maintains a set of 8 associativity change counters in
1147 * the VPA of each cpu that correspond to the associativity levels in the
1148 * ibm,associativity-reference-points property. When an associativity
1149 * level changes, the corresponding counter is incremented.
1151 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1152 * node associativity levels have changed.
1154 * Returns the number of cpus with unhandled associativity changes.
1156 static int update_cpu_associativity_changes_mask(void)
1159 cpumask_t *changes = &cpu_associativity_changes_mask;
1161 for_each_possible_cpu(cpu) {
1163 u8 *counts = vphn_cpu_change_counts[cpu];
1164 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1166 for (i = 0; i < distance_ref_points_depth; i++) {
1167 if (hypervisor_counts[i] != counts[i]) {
1168 counts[i] = hypervisor_counts[i];
1173 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1174 cpu = cpu_last_thread_sibling(cpu);
1178 return cpumask_weight(changes);
1182 * Retrieve the new associativity information for a virtual processor's
1185 static long vphn_get_associativity(unsigned long cpu,
1186 __be32 *associativity)
1190 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1191 VPHN_FLAG_VCPU, associativity);
1195 printk_once(KERN_INFO
1196 "VPHN is not supported. Disabling polling...\n");
1197 stop_topology_update();
1201 "hcall_vphn() experienced a hardware fault "
1202 "preventing VPHN. Disabling polling...\n");
1203 stop_topology_update();
1206 dbg("VPHN hcall succeeded. Reset polling...\n");
1207 timed_topology_update(0);
1214 int find_and_online_cpu_nid(int cpu)
1216 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1219 /* Use associativity from first thread for all siblings */
1220 if (vphn_get_associativity(cpu, associativity))
1221 return cpu_to_node(cpu);
1223 new_nid = associativity_to_nid(associativity);
1224 if (new_nid < 0 || !node_possible(new_nid))
1225 new_nid = first_online_node;
1227 if (NODE_DATA(new_nid) == NULL) {
1228 #ifdef CONFIG_MEMORY_HOTPLUG
1230 * Need to ensure that NODE_DATA is initialized for a node from
1231 * available memory (see memblock_alloc_try_nid). If unable to
1232 * init the node, then default to nearest node that has memory
1233 * installed. Skip onlining a node if the subsystems are not
1236 if (!topology_inited || try_online_node(new_nid))
1237 new_nid = first_online_node;
1240 * Default to using the nearest node that has memory installed.
1241 * Otherwise, it would be necessary to patch the kernel MM code
1242 * to deal with more memoryless-node error conditions.
1244 new_nid = first_online_node;
1248 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1254 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1255 * characteristics change. This function doesn't perform any locking and is
1256 * only safe to call from stop_machine().
1258 static int update_cpu_topology(void *data)
1260 struct topology_update_data *update;
1266 cpu = smp_processor_id();
1268 for (update = data; update; update = update->next) {
1269 int new_nid = update->new_nid;
1270 if (cpu != update->cpu)
1273 unmap_cpu_from_node(cpu);
1274 map_cpu_to_node(cpu, new_nid);
1275 set_cpu_numa_node(cpu, new_nid);
1276 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1283 static int update_lookup_table(void *data)
1285 struct topology_update_data *update;
1291 * Upon topology update, the numa-cpu lookup table needs to be updated
1292 * for all threads in the core, including offline CPUs, to ensure that
1293 * future hotplug operations respect the cpu-to-node associativity
1296 for (update = data; update; update = update->next) {
1299 nid = update->new_nid;
1300 base = cpu_first_thread_sibling(update->cpu);
1302 for (j = 0; j < threads_per_core; j++) {
1303 update_numa_cpu_lookup_table(base + j, nid);
1311 * Update the node maps and sysfs entries for each cpu whose home node
1312 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1314 * cpus_locked says whether we already hold cpu_hotplug_lock.
1316 int numa_update_cpu_topology(bool cpus_locked)
1318 unsigned int cpu, sibling, changed = 0;
1319 struct topology_update_data *updates, *ud;
1320 cpumask_t updated_cpus;
1322 int weight, new_nid, i = 0;
1324 if (!prrn_enabled && !vphn_enabled && topology_inited)
1327 weight = cpumask_weight(&cpu_associativity_changes_mask);
1331 updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL);
1335 cpumask_clear(&updated_cpus);
1337 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1339 * If siblings aren't flagged for changes, updates list
1340 * will be too short. Skip on this update and set for next
1343 if (!cpumask_subset(cpu_sibling_mask(cpu),
1344 &cpu_associativity_changes_mask)) {
1345 pr_info("Sibling bits not set for associativity "
1346 "change, cpu%d\n", cpu);
1347 cpumask_or(&cpu_associativity_changes_mask,
1348 &cpu_associativity_changes_mask,
1349 cpu_sibling_mask(cpu));
1350 cpu = cpu_last_thread_sibling(cpu);
1354 new_nid = find_and_online_cpu_nid(cpu);
1356 if (new_nid == numa_cpu_lookup_table[cpu]) {
1357 cpumask_andnot(&cpu_associativity_changes_mask,
1358 &cpu_associativity_changes_mask,
1359 cpu_sibling_mask(cpu));
1360 dbg("Assoc chg gives same node %d for cpu%d\n",
1362 cpu = cpu_last_thread_sibling(cpu);
1366 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1368 ud->next = &updates[i];
1370 ud->new_nid = new_nid;
1371 ud->old_nid = numa_cpu_lookup_table[sibling];
1372 cpumask_set_cpu(sibling, &updated_cpus);
1374 cpu = cpu_last_thread_sibling(cpu);
1378 * Prevent processing of 'updates' from overflowing array
1379 * where last entry filled in a 'next' pointer.
1382 updates[i-1].next = NULL;
1384 pr_debug("Topology update for the following CPUs:\n");
1385 if (cpumask_weight(&updated_cpus)) {
1386 for (ud = &updates[0]; ud; ud = ud->next) {
1387 pr_debug("cpu %d moving from node %d "
1389 ud->old_nid, ud->new_nid);
1394 * In cases where we have nothing to update (because the updates list
1395 * is too short or because the new topology is same as the old one),
1396 * skip invoking update_cpu_topology() via stop-machine(). This is
1397 * necessary (and not just a fast-path optimization) since stop-machine
1398 * can end up electing a random CPU to run update_cpu_topology(), and
1399 * thus trick us into setting up incorrect cpu-node mappings (since
1400 * 'updates' is kzalloc()'ed).
1402 * And for the similar reason, we will skip all the following updating.
1404 if (!cpumask_weight(&updated_cpus))
1408 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1411 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1414 * Update the numa-cpu lookup table with the new mappings, even for
1415 * offline CPUs. It is best to perform this update from the stop-
1419 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1420 cpumask_of(raw_smp_processor_id()));
1422 stop_machine(update_lookup_table, &updates[0],
1423 cpumask_of(raw_smp_processor_id()));
1425 for (ud = &updates[0]; ud; ud = ud->next) {
1426 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1427 register_cpu_under_node(ud->cpu, ud->new_nid);
1429 dev = get_cpu_device(ud->cpu);
1431 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1432 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1441 int arch_update_cpu_topology(void)
1443 return numa_update_cpu_topology(true);
1446 static void topology_work_fn(struct work_struct *work)
1448 rebuild_sched_domains();
1450 static DECLARE_WORK(topology_work, topology_work_fn);
1452 static void topology_schedule_update(void)
1454 schedule_work(&topology_work);
1457 static void topology_timer_fn(struct timer_list *unused)
1459 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1460 topology_schedule_update();
1461 else if (vphn_enabled) {
1462 if (update_cpu_associativity_changes_mask() > 0)
1463 topology_schedule_update();
1464 reset_topology_timer();
1467 static struct timer_list topology_timer;
1469 static void reset_topology_timer(void)
1472 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1477 static int dt_update_callback(struct notifier_block *nb,
1478 unsigned long action, void *data)
1480 struct of_reconfig_data *update = data;
1481 int rc = NOTIFY_DONE;
1484 case OF_RECONFIG_UPDATE_PROPERTY:
1485 if (of_node_is_type(update->dn, "cpu") &&
1486 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1488 of_property_read_u32(update->dn, "reg", &core_id);
1489 rc = dlpar_cpu_readd(core_id);
1498 static struct notifier_block dt_update_nb = {
1499 .notifier_call = dt_update_callback,
1505 * Start polling for associativity changes.
1507 int start_topology_update(void)
1511 if (!topology_updates_enabled)
1514 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1515 if (!prrn_enabled) {
1518 rc = of_reconfig_notifier_register(&dt_update_nb);
1522 if (firmware_has_feature(FW_FEATURE_VPHN) &&
1523 lppaca_shared_proc(get_lppaca())) {
1524 if (!vphn_enabled) {
1526 setup_cpu_associativity_change_counters();
1527 timer_setup(&topology_timer, topology_timer_fn,
1529 reset_topology_timer();
1533 pr_info("Starting topology update%s%s\n",
1534 (prrn_enabled ? " prrn_enabled" : ""),
1535 (vphn_enabled ? " vphn_enabled" : ""));
1541 * Disable polling for VPHN associativity changes.
1543 int stop_topology_update(void)
1547 if (!topology_updates_enabled)
1553 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1558 rc = del_timer_sync(&topology_timer);
1561 pr_info("Stopping topology update\n");
1566 int prrn_is_enabled(void)
1568 return prrn_enabled;
1571 void __init shared_proc_topology_init(void)
1573 if (lppaca_shared_proc(get_lppaca())) {
1574 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1576 numa_update_cpu_topology(false);
1580 static int topology_read(struct seq_file *file, void *v)
1582 if (vphn_enabled || prrn_enabled)
1583 seq_puts(file, "on\n");
1585 seq_puts(file, "off\n");
1590 static int topology_open(struct inode *inode, struct file *file)
1592 return single_open(file, topology_read, NULL);
1595 static ssize_t topology_write(struct file *file, const char __user *buf,
1596 size_t count, loff_t *off)
1598 char kbuf[4]; /* "on" or "off" plus null. */
1601 read_len = count < 3 ? count : 3;
1602 if (copy_from_user(kbuf, buf, read_len))
1605 kbuf[read_len] = '\0';
1607 if (!strncmp(kbuf, "on", 2)) {
1608 topology_updates_enabled = true;
1609 start_topology_update();
1610 } else if (!strncmp(kbuf, "off", 3)) {
1611 stop_topology_update();
1612 topology_updates_enabled = false;
1619 static const struct file_operations topology_ops = {
1621 .write = topology_write,
1622 .open = topology_open,
1623 .release = single_release
1626 static int topology_update_init(void)
1628 start_topology_update();
1631 topology_schedule_update();
1633 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1636 topology_inited = 1;
1639 device_initcall(topology_update_init);
1640 #endif /* CONFIG_PPC_SPLPAR */