4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
11 #define pr_fmt(fmt) "numa: " fmt
13 #include <linux/threads.h>
14 #include <linux/bootmem.h>
15 #include <linux/init.h>
17 #include <linux/mmzone.h>
18 #include <linux/export.h>
19 #include <linux/nodemask.h>
20 #include <linux/cpu.h>
21 #include <linux/notifier.h>
22 #include <linux/memblock.h>
24 #include <linux/pfn.h>
25 #include <linux/cpuset.h>
26 #include <linux/node.h>
27 #include <linux/stop_machine.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/uaccess.h>
31 #include <linux/slab.h>
32 #include <asm/cputhreads.h>
33 #include <asm/sparsemem.h>
36 #include <asm/cputhreads.h>
37 #include <asm/topology.h>
38 #include <asm/firmware.h>
40 #include <asm/hvcall.h>
41 #include <asm/setup.h>
44 static int numa_enabled = 1;
46 static char *cmdline __initdata;
48 static int numa_debug;
49 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
51 int numa_cpu_lookup_table[NR_CPUS];
52 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
53 struct pglist_data *node_data[MAX_NUMNODES];
55 EXPORT_SYMBOL(numa_cpu_lookup_table);
56 EXPORT_SYMBOL(node_to_cpumask_map);
57 EXPORT_SYMBOL(node_data);
59 static int min_common_depth;
60 static int n_mem_addr_cells, n_mem_size_cells;
61 static int form1_affinity;
63 #define MAX_DISTANCE_REF_POINTS 4
64 static int distance_ref_points_depth;
65 static const __be32 *distance_ref_points;
66 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
69 * Allocate node_to_cpumask_map based on number of available nodes
70 * Requires node_possible_map to be valid.
72 * Note: cpumask_of_node() is not valid until after this is done.
74 static void __init setup_node_to_cpumask_map(void)
78 /* setup nr_node_ids if not done yet */
79 if (nr_node_ids == MAX_NUMNODES)
82 /* allocate the map */
84 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
86 /* cpumask_of_node() will now work */
87 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
90 static int __init fake_numa_create_new_node(unsigned long end_pfn,
93 unsigned long long mem;
95 static unsigned int fake_nid;
96 static unsigned long long curr_boundary;
99 * Modify node id, iff we started creating NUMA nodes
100 * We want to continue from where we left of the last time
105 * In case there are no more arguments to parse, the
106 * node_id should be the same as the last fake node id
107 * (we've handled this above).
112 mem = memparse(p, &p);
116 if (mem < curr_boundary)
121 if ((end_pfn << PAGE_SHIFT) > mem) {
123 * Skip commas and spaces
125 while (*p == ',' || *p == ' ' || *p == '\t')
131 dbg("created new fake_node with id %d\n", fake_nid);
137 static void reset_numa_cpu_lookup_table(void)
141 for_each_possible_cpu(cpu)
142 numa_cpu_lookup_table[cpu] = -1;
145 static void map_cpu_to_node(int cpu, int node)
147 update_numa_cpu_lookup_table(cpu, node);
149 dbg("adding cpu %d to node %d\n", cpu, node);
151 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
152 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
155 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
156 static void unmap_cpu_from_node(unsigned long cpu)
158 int node = numa_cpu_lookup_table[cpu];
160 dbg("removing cpu %lu from node %d\n", cpu, node);
162 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
163 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
165 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
169 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
171 /* must hold reference to node during call */
172 static const __be32 *of_get_associativity(struct device_node *dev)
174 return of_get_property(dev, "ibm,associativity", NULL);
178 * Returns the property linux,drconf-usable-memory if
179 * it exists (the property exists only in kexec/kdump kernels,
180 * added by kexec-tools)
182 static const __be32 *of_get_usable_memory(struct device_node *memory)
186 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
187 if (!prop || len < sizeof(unsigned int))
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)
235 if (min_common_depth == -1)
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)
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)
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);
386 * Read the next memblock list entry from the ibm,dynamic-memory property
387 * and return the information in the provided of_drconf_cell structure.
389 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
393 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
396 drmem->drc_index = of_read_number(cp, 1);
397 drmem->reserved = of_read_number(&cp[1], 1);
398 drmem->aa_index = of_read_number(&cp[2], 1);
399 drmem->flags = of_read_number(&cp[3], 1);
405 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
407 * The layout of the ibm,dynamic-memory property is a number N of memblock
408 * list entries followed by N memblock list entries. Each memblock list entry
409 * contains information as laid out in the of_drconf_cell struct above.
411 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
416 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
417 if (!prop || len < sizeof(unsigned int))
420 entries = of_read_number(prop++, 1);
422 /* Now that we know the number of entries, revalidate the size
423 * of the property read in to ensure we have everything
425 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
433 * Retrieve and validate the ibm,lmb-size property for drconf memory
434 * from the device tree.
436 static u64 of_get_lmb_size(struct device_node *memory)
441 prop = of_get_property(memory, "ibm,lmb-size", &len);
442 if (!prop || len < sizeof(unsigned int))
445 return read_n_cells(n_mem_size_cells, &prop);
448 struct assoc_arrays {
451 const __be32 *arrays;
455 * Retrieve and validate the list of associativity arrays for drconf
456 * memory from the ibm,associativity-lookup-arrays property of the
459 * The layout of the ibm,associativity-lookup-arrays property is a number N
460 * indicating the number of associativity arrays, followed by a number M
461 * indicating the size of each associativity array, followed by a list
462 * of N associativity arrays.
464 static int of_get_assoc_arrays(struct device_node *memory,
465 struct assoc_arrays *aa)
470 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
471 if (!prop || len < 2 * sizeof(unsigned int))
474 aa->n_arrays = of_read_number(prop++, 1);
475 aa->array_sz = of_read_number(prop++, 1);
477 /* Now that we know the number of arrays and size of each array,
478 * revalidate the size of the property read in.
480 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
488 * This is like of_node_to_nid_single() for memory represented in the
489 * ibm,dynamic-reconfiguration-memory node.
491 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
492 struct assoc_arrays *aa)
495 int nid = default_nid;
498 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
499 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
500 drmem->aa_index < aa->n_arrays) {
501 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
502 nid = of_read_number(&aa->arrays[index], 1);
504 if (nid == 0xffff || nid >= MAX_NUMNODES)
508 index = drmem->aa_index * aa->array_sz;
509 initialize_distance_lookup_table(nid,
518 * Figure out to which domain a cpu belongs and stick it there.
519 * Return the id of the domain used.
521 static int numa_setup_cpu(unsigned long lcpu)
524 struct device_node *cpu;
527 * If a valid cpu-to-node mapping is already available, use it
528 * directly instead of querying the firmware, since it represents
529 * the most recent mapping notified to us by the platform (eg: VPHN).
531 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
532 map_cpu_to_node(lcpu, nid);
536 cpu = of_get_cpu_node(lcpu, NULL);
540 if (cpu_present(lcpu))
546 nid = of_node_to_nid_single(cpu);
549 if (nid < 0 || !node_possible(nid))
550 nid = first_online_node;
552 map_cpu_to_node(lcpu, nid);
558 static void verify_cpu_node_mapping(int cpu, int node)
560 int base, sibling, i;
562 /* Verify that all the threads in the core belong to the same node */
563 base = cpu_first_thread_sibling(cpu);
565 for (i = 0; i < threads_per_core; i++) {
568 if (sibling == cpu || cpu_is_offline(sibling))
571 if (cpu_to_node(sibling) != node) {
572 WARN(1, "CPU thread siblings %d and %d don't belong"
573 " to the same node!\n", cpu, sibling);
579 /* Must run before sched domains notifier. */
580 static int ppc_numa_cpu_prepare(unsigned int cpu)
584 nid = numa_setup_cpu(cpu);
585 verify_cpu_node_mapping(cpu, nid);
589 static int ppc_numa_cpu_dead(unsigned int cpu)
591 #ifdef CONFIG_HOTPLUG_CPU
592 unmap_cpu_from_node(cpu);
598 * Check and possibly modify a memory region to enforce the memory limit.
600 * Returns the size the region should have to enforce the memory limit.
601 * This will either be the original value of size, a truncated value,
602 * or zero. If the returned value of size is 0 the region should be
603 * discarded as it lies wholly above the memory limit.
605 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
609 * We use memblock_end_of_DRAM() in here instead of memory_limit because
610 * we've already adjusted it for the limit and it takes care of
611 * having memory holes below the limit. Also, in the case of
612 * iommu_is_off, memory_limit is not set but is implicitly enforced.
615 if (start + size <= memblock_end_of_DRAM())
618 if (start >= memblock_end_of_DRAM())
621 return memblock_end_of_DRAM() - start;
625 * Reads the counter for a given entry in
626 * linux,drconf-usable-memory property
628 static inline int __init read_usm_ranges(const __be32 **usm)
631 * For each lmb in ibm,dynamic-memory a corresponding
632 * entry in linux,drconf-usable-memory property contains
633 * a counter followed by that many (base, size) duple.
634 * read the counter from linux,drconf-usable-memory
636 return read_n_cells(n_mem_size_cells, usm);
640 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
641 * node. This assumes n_mem_{addr,size}_cells have been set.
643 static void __init parse_drconf_memory(struct device_node *memory)
645 const __be32 *uninitialized_var(dm), *usm;
646 unsigned int n, rc, ranges, is_kexec_kdump = 0;
647 unsigned long lmb_size, base, size, sz;
649 struct assoc_arrays aa = { .arrays = NULL };
651 n = of_get_drconf_memory(memory, &dm);
655 lmb_size = of_get_lmb_size(memory);
659 rc = of_get_assoc_arrays(memory, &aa);
663 /* check if this is a kexec/kdump kernel */
664 usm = of_get_usable_memory(memory);
668 for (; n != 0; --n) {
669 struct of_drconf_cell drmem;
671 read_drconf_cell(&drmem, &dm);
673 /* skip this block if the reserved bit is set in flags (0x80)
674 or if the block is not assigned to this partition (0x8) */
675 if ((drmem.flags & DRCONF_MEM_RESERVED)
676 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
679 base = drmem.base_addr;
683 if (is_kexec_kdump) {
684 ranges = read_usm_ranges(&usm);
685 if (!ranges) /* there are no (base, size) duple */
689 if (is_kexec_kdump) {
690 base = read_n_cells(n_mem_addr_cells, &usm);
691 size = read_n_cells(n_mem_size_cells, &usm);
693 nid = of_drconf_to_nid_single(&drmem, &aa);
694 fake_numa_create_new_node(
695 ((base + size) >> PAGE_SHIFT),
697 node_set_online(nid);
698 sz = numa_enforce_memory_limit(base, size);
700 memblock_set_node(base, sz,
701 &memblock.memory, nid);
706 static int __init parse_numa_properties(void)
708 struct device_node *memory;
712 if (numa_enabled == 0) {
713 printk(KERN_WARNING "NUMA disabled by user\n");
717 min_common_depth = find_min_common_depth();
719 if (min_common_depth < 0)
720 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;
733 cpu = of_get_cpu_node(i, NULL);
735 nid = of_node_to_nid_single(cpu);
739 * Don't fall back to default_nid yet -- we will plug
740 * cpus into nodes once the memory scan has discovered
745 node_set_online(nid);
748 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
750 for_each_node_by_type(memory, "memory") {
755 const __be32 *memcell_buf;
758 memcell_buf = of_get_property(memory,
759 "linux,usable-memory", &len);
760 if (!memcell_buf || len <= 0)
761 memcell_buf = of_get_property(memory, "reg", &len);
762 if (!memcell_buf || len <= 0)
766 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
768 /* these are order-sensitive, and modify the buffer pointer */
769 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
770 size = read_n_cells(n_mem_size_cells, &memcell_buf);
773 * Assumption: either all memory nodes or none will
774 * have associativity properties. If none, then
775 * everything goes to default_nid.
777 nid = of_node_to_nid_single(memory);
781 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
782 node_set_online(nid);
784 size = numa_enforce_memory_limit(start, size);
786 memblock_set_node(start, size, &memblock.memory, nid);
793 * Now do the same thing for each MEMBLOCK listed in the
794 * ibm,dynamic-memory property in the
795 * ibm,dynamic-reconfiguration-memory node.
797 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
799 parse_drconf_memory(memory);
804 static void __init setup_nonnuma(void)
806 unsigned long top_of_ram = memblock_end_of_DRAM();
807 unsigned long total_ram = memblock_phys_mem_size();
808 unsigned long start_pfn, end_pfn;
809 unsigned int nid = 0;
810 struct memblock_region *reg;
812 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
813 top_of_ram, total_ram);
814 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
815 (top_of_ram - total_ram) >> 20);
817 for_each_memblock(memory, reg) {
818 start_pfn = memblock_region_memory_base_pfn(reg);
819 end_pfn = memblock_region_memory_end_pfn(reg);
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;
834 if (min_common_depth == -1 || !numa_enabled)
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_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
876 /* report and initialize */
877 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
878 nd_pa, nd_pa + nd_size - 1);
879 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
881 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
884 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
885 NODE_DATA(nid)->node_id = nid;
886 NODE_DATA(nid)->node_start_pfn = start_pfn;
887 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
890 static void __init find_possible_nodes(void)
892 struct device_node *rtas;
895 if (min_common_depth <= 0)
898 rtas = of_find_node_by_path("/rtas");
902 if (of_property_read_u32_index(rtas,
903 "ibm,max-associativity-domains",
904 min_common_depth, &numnodes))
907 for (i = 0; i < numnodes; i++) {
908 if (!node_possible(i))
909 node_set(i, node_possible_map);
916 void __init initmem_init(void)
920 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
921 max_pfn = max_low_pfn;
923 if (parse_numa_properties())
929 * Modify the set of possible NUMA nodes to reflect information
930 * available about the set of online nodes, and the set of nodes
931 * that we expect to make use of for this platform's affinity
934 nodes_and(node_possible_map, node_possible_map, node_online_map);
936 find_possible_nodes();
938 for_each_online_node(nid) {
939 unsigned long start_pfn, end_pfn;
941 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
942 setup_node_data(nid, start_pfn, end_pfn);
943 sparse_memory_present_with_active_regions(nid);
948 setup_node_to_cpumask_map();
950 reset_numa_cpu_lookup_table();
953 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
954 * even before we online them, so that we can use cpu_to_{node,mem}
955 * early in boot, cf. smp_prepare_cpus().
956 * _nocalls() + manual invocation is used because cpuhp is not yet
957 * initialized for the boot CPU.
959 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
960 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
961 for_each_present_cpu(cpu)
965 static int __init early_numa(char *p)
970 if (strstr(p, "off"))
973 if (strstr(p, "debug"))
976 p = strstr(p, "fake=");
978 cmdline = p + strlen("fake=");
982 early_param("numa", early_numa);
984 static bool topology_updates_enabled = true;
986 static int __init early_topology_updates(char *p)
991 if (!strcmp(p, "off")) {
992 pr_info("Disabling topology updates\n");
993 topology_updates_enabled = false;
998 early_param("topology_updates", early_topology_updates);
1000 #ifdef CONFIG_MEMORY_HOTPLUG
1002 * Find the node associated with a hot added memory section for
1003 * memory represented in the device tree by the property
1004 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1006 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1007 unsigned long scn_addr)
1010 unsigned int drconf_cell_cnt, rc;
1011 unsigned long lmb_size;
1012 struct assoc_arrays aa;
1015 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1016 if (!drconf_cell_cnt)
1019 lmb_size = of_get_lmb_size(memory);
1023 rc = of_get_assoc_arrays(memory, &aa);
1027 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1028 struct of_drconf_cell drmem;
1030 read_drconf_cell(&drmem, &dm);
1032 /* skip this block if it is reserved or not assigned to
1034 if ((drmem.flags & DRCONF_MEM_RESERVED)
1035 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1038 if ((scn_addr < drmem.base_addr)
1039 || (scn_addr >= (drmem.base_addr + lmb_size)))
1042 nid = of_drconf_to_nid_single(&drmem, &aa);
1050 * Find the node associated with a hot added memory section for memory
1051 * represented in the device tree as a node (i.e. memory@XXXX) for
1054 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1056 struct device_node *memory;
1059 for_each_node_by_type(memory, "memory") {
1060 unsigned long start, size;
1062 const __be32 *memcell_buf;
1065 memcell_buf = of_get_property(memory, "reg", &len);
1066 if (!memcell_buf || len <= 0)
1069 /* ranges in cell */
1070 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1073 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1074 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1076 if ((scn_addr < start) || (scn_addr >= (start + size)))
1079 nid = of_node_to_nid_single(memory);
1087 of_node_put(memory);
1093 * Find the node associated with a hot added memory section. Section
1094 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1095 * sections are fully contained within a single MEMBLOCK.
1097 int hot_add_scn_to_nid(unsigned long scn_addr)
1099 struct device_node *memory = NULL;
1102 if (!numa_enabled || (min_common_depth < 0))
1103 return first_online_node;
1105 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1107 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1108 of_node_put(memory);
1110 nid = hot_add_node_scn_to_nid(scn_addr);
1113 if (nid < 0 || !node_possible(nid))
1114 nid = first_online_node;
1119 static u64 hot_add_drconf_memory_max(void)
1121 struct device_node *memory = NULL;
1122 struct device_node *dn = NULL;
1123 unsigned int drconf_cell_cnt = 0;
1125 const __be32 *dm = NULL;
1126 const __be64 *lrdr = NULL;
1127 struct of_drconf_cell drmem;
1129 dn = of_find_node_by_path("/rtas");
1131 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1134 return be64_to_cpup(lrdr);
1137 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1139 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1140 lmb_size = of_get_lmb_size(memory);
1142 /* Advance to the last cell, each cell has 6 32 bit integers */
1143 dm += (drconf_cell_cnt - 1) * 6;
1144 read_drconf_cell(&drmem, &dm);
1145 of_node_put(memory);
1146 return drmem.base_addr + lmb_size;
1152 * memory_hotplug_max - return max address of memory that may be added
1154 * This is currently only used on systems that support drconfig memory
1157 u64 memory_hotplug_max(void)
1159 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1161 #endif /* CONFIG_MEMORY_HOTPLUG */
1163 /* Virtual Processor Home Node (VPHN) support */
1164 #ifdef CONFIG_PPC_SPLPAR
1168 struct topology_update_data {
1169 struct topology_update_data *next;
1175 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1176 static cpumask_t cpu_associativity_changes_mask;
1177 static int vphn_enabled;
1178 static int prrn_enabled;
1179 static void reset_topology_timer(void);
1182 * Store the current values of the associativity change counters in the
1185 static void setup_cpu_associativity_change_counters(void)
1189 /* The VPHN feature supports a maximum of 8 reference points */
1190 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1192 for_each_possible_cpu(cpu) {
1194 u8 *counts = vphn_cpu_change_counts[cpu];
1195 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1197 for (i = 0; i < distance_ref_points_depth; i++)
1198 counts[i] = hypervisor_counts[i];
1203 * The hypervisor maintains a set of 8 associativity change counters in
1204 * the VPA of each cpu that correspond to the associativity levels in the
1205 * ibm,associativity-reference-points property. When an associativity
1206 * level changes, the corresponding counter is incremented.
1208 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1209 * node associativity levels have changed.
1211 * Returns the number of cpus with unhandled associativity changes.
1213 static int update_cpu_associativity_changes_mask(void)
1216 cpumask_t *changes = &cpu_associativity_changes_mask;
1218 for_each_possible_cpu(cpu) {
1220 u8 *counts = vphn_cpu_change_counts[cpu];
1221 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1223 for (i = 0; i < distance_ref_points_depth; i++) {
1224 if (hypervisor_counts[i] != counts[i]) {
1225 counts[i] = hypervisor_counts[i];
1230 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1231 cpu = cpu_last_thread_sibling(cpu);
1235 return cpumask_weight(changes);
1239 * Retrieve the new associativity information for a virtual processor's
1242 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1245 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1247 int hwcpu = get_hard_smp_processor_id(cpu);
1249 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1250 vphn_unpack_associativity(retbuf, associativity);
1255 static long vphn_get_associativity(unsigned long cpu,
1256 __be32 *associativity)
1260 rc = hcall_vphn(cpu, associativity);
1264 printk_once(KERN_INFO
1265 "VPHN is not supported. Disabling polling...\n");
1266 stop_topology_update();
1270 "hcall_vphn() experienced a hardware fault "
1271 "preventing VPHN. Disabling polling...\n");
1272 stop_topology_update();
1278 static inline int find_and_online_cpu_nid(int cpu)
1280 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1283 /* Use associativity from first thread for all siblings */
1284 vphn_get_associativity(cpu, associativity);
1285 new_nid = associativity_to_nid(associativity);
1286 if (new_nid < 0 || !node_possible(new_nid))
1287 new_nid = first_online_node;
1289 if (NODE_DATA(new_nid) == NULL) {
1290 #ifdef CONFIG_MEMORY_HOTPLUG
1292 * Need to ensure that NODE_DATA is initialized for a node from
1293 * available memory (see memblock_alloc_try_nid). If unable to
1294 * init the node, then default to nearest node that has memory
1297 if (try_online_node(new_nid))
1298 new_nid = first_online_node;
1301 * Default to using the nearest node that has memory installed.
1302 * Otherwise, it would be necessary to patch the kernel MM code
1303 * to deal with more memoryless-node error conditions.
1305 new_nid = first_online_node;
1313 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1314 * characteristics change. This function doesn't perform any locking and is
1315 * only safe to call from stop_machine().
1317 static int update_cpu_topology(void *data)
1319 struct topology_update_data *update;
1325 cpu = smp_processor_id();
1327 for (update = data; update; update = update->next) {
1328 int new_nid = update->new_nid;
1329 if (cpu != update->cpu)
1332 unmap_cpu_from_node(cpu);
1333 map_cpu_to_node(cpu, new_nid);
1334 set_cpu_numa_node(cpu, new_nid);
1335 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1342 static int update_lookup_table(void *data)
1344 struct topology_update_data *update;
1350 * Upon topology update, the numa-cpu lookup table needs to be updated
1351 * for all threads in the core, including offline CPUs, to ensure that
1352 * future hotplug operations respect the cpu-to-node associativity
1355 for (update = data; update; update = update->next) {
1358 nid = update->new_nid;
1359 base = cpu_first_thread_sibling(update->cpu);
1361 for (j = 0; j < threads_per_core; j++) {
1362 update_numa_cpu_lookup_table(base + j, nid);
1370 * Update the node maps and sysfs entries for each cpu whose home node
1371 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1373 * cpus_locked says whether we already hold cpu_hotplug_lock.
1375 int numa_update_cpu_topology(bool cpus_locked)
1377 unsigned int cpu, sibling, changed = 0;
1378 struct topology_update_data *updates, *ud;
1379 cpumask_t updated_cpus;
1381 int weight, new_nid, i = 0;
1383 if (!prrn_enabled && !vphn_enabled)
1386 weight = cpumask_weight(&cpu_associativity_changes_mask);
1390 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1394 cpumask_clear(&updated_cpus);
1396 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1398 * If siblings aren't flagged for changes, updates list
1399 * will be too short. Skip on this update and set for next
1402 if (!cpumask_subset(cpu_sibling_mask(cpu),
1403 &cpu_associativity_changes_mask)) {
1404 pr_info("Sibling bits not set for associativity "
1405 "change, cpu%d\n", cpu);
1406 cpumask_or(&cpu_associativity_changes_mask,
1407 &cpu_associativity_changes_mask,
1408 cpu_sibling_mask(cpu));
1409 cpu = cpu_last_thread_sibling(cpu);
1413 new_nid = find_and_online_cpu_nid(cpu);
1415 if (new_nid == numa_cpu_lookup_table[cpu]) {
1416 cpumask_andnot(&cpu_associativity_changes_mask,
1417 &cpu_associativity_changes_mask,
1418 cpu_sibling_mask(cpu));
1419 cpu = cpu_last_thread_sibling(cpu);
1423 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1426 ud->new_nid = new_nid;
1427 ud->old_nid = numa_cpu_lookup_table[sibling];
1428 cpumask_set_cpu(sibling, &updated_cpus);
1430 ud->next = &updates[i];
1432 cpu = cpu_last_thread_sibling(cpu);
1435 pr_debug("Topology update for the following CPUs:\n");
1436 if (cpumask_weight(&updated_cpus)) {
1437 for (ud = &updates[0]; ud; ud = ud->next) {
1438 pr_debug("cpu %d moving from node %d "
1440 ud->old_nid, ud->new_nid);
1445 * In cases where we have nothing to update (because the updates list
1446 * is too short or because the new topology is same as the old one),
1447 * skip invoking update_cpu_topology() via stop-machine(). This is
1448 * necessary (and not just a fast-path optimization) since stop-machine
1449 * can end up electing a random CPU to run update_cpu_topology(), and
1450 * thus trick us into setting up incorrect cpu-node mappings (since
1451 * 'updates' is kzalloc()'ed).
1453 * And for the similar reason, we will skip all the following updating.
1455 if (!cpumask_weight(&updated_cpus))
1459 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1462 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1465 * Update the numa-cpu lookup table with the new mappings, even for
1466 * offline CPUs. It is best to perform this update from the stop-
1470 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1471 cpumask_of(raw_smp_processor_id()));
1473 stop_machine(update_lookup_table, &updates[0],
1474 cpumask_of(raw_smp_processor_id()));
1476 for (ud = &updates[0]; ud; ud = ud->next) {
1477 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1478 register_cpu_under_node(ud->cpu, ud->new_nid);
1480 dev = get_cpu_device(ud->cpu);
1482 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1483 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1492 int arch_update_cpu_topology(void)
1494 return numa_update_cpu_topology(true);
1497 static void topology_work_fn(struct work_struct *work)
1499 rebuild_sched_domains();
1501 static DECLARE_WORK(topology_work, topology_work_fn);
1503 static void topology_schedule_update(void)
1505 schedule_work(&topology_work);
1508 static void topology_timer_fn(unsigned long ignored)
1510 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1511 topology_schedule_update();
1512 else if (vphn_enabled) {
1513 if (update_cpu_associativity_changes_mask() > 0)
1514 topology_schedule_update();
1515 reset_topology_timer();
1518 static struct timer_list topology_timer =
1519 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1521 static void reset_topology_timer(void)
1523 topology_timer.data = 0;
1524 topology_timer.expires = jiffies + 60 * HZ;
1525 mod_timer(&topology_timer, topology_timer.expires);
1530 static int dt_update_callback(struct notifier_block *nb,
1531 unsigned long action, void *data)
1533 struct of_reconfig_data *update = data;
1534 int rc = NOTIFY_DONE;
1537 case OF_RECONFIG_UPDATE_PROPERTY:
1538 if (!of_prop_cmp(update->dn->type, "cpu") &&
1539 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1541 of_property_read_u32(update->dn, "reg", &core_id);
1542 rc = dlpar_cpu_readd(core_id);
1551 static struct notifier_block dt_update_nb = {
1552 .notifier_call = dt_update_callback,
1558 * Start polling for associativity changes.
1560 int start_topology_update(void)
1564 if (!topology_updates_enabled)
1567 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1568 if (!prrn_enabled) {
1572 rc = of_reconfig_notifier_register(&dt_update_nb);
1575 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1576 lppaca_shared_proc(get_lppaca())) {
1577 if (!vphn_enabled) {
1580 setup_cpu_associativity_change_counters();
1581 init_timer_deferrable(&topology_timer);
1582 reset_topology_timer();
1590 * Disable polling for VPHN associativity changes.
1592 int stop_topology_update(void)
1596 if (!topology_updates_enabled)
1602 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1604 } else if (vphn_enabled) {
1606 rc = del_timer_sync(&topology_timer);
1612 int prrn_is_enabled(void)
1614 return prrn_enabled;
1617 static int topology_read(struct seq_file *file, void *v)
1619 if (vphn_enabled || prrn_enabled)
1620 seq_puts(file, "on\n");
1622 seq_puts(file, "off\n");
1627 static int topology_open(struct inode *inode, struct file *file)
1629 return single_open(file, topology_read, NULL);
1632 static ssize_t topology_write(struct file *file, const char __user *buf,
1633 size_t count, loff_t *off)
1635 char kbuf[4]; /* "on" or "off" plus null. */
1638 read_len = count < 3 ? count : 3;
1639 if (copy_from_user(kbuf, buf, read_len))
1642 kbuf[read_len] = '\0';
1644 if (!strncmp(kbuf, "on", 2)) {
1645 topology_updates_enabled = true;
1646 start_topology_update();
1647 } else if (!strncmp(kbuf, "off", 3)) {
1648 stop_topology_update();
1649 topology_updates_enabled = false;
1656 static const struct file_operations topology_ops = {
1658 .write = topology_write,
1659 .open = topology_open,
1660 .release = single_release
1663 static int topology_update_init(void)
1665 start_topology_update();
1667 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1672 device_initcall(topology_update_init);
1673 #endif /* CONFIG_PPC_SPLPAR */