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 update_numa_cpu_lookup_table(unsigned int cpu, int node)
147 numa_cpu_lookup_table[cpu] = node;
150 static void map_cpu_to_node(int cpu, int node)
152 update_numa_cpu_lookup_table(cpu, node);
154 dbg("adding cpu %d to node %d\n", cpu, node);
156 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
157 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
160 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
161 static void unmap_cpu_from_node(unsigned long cpu)
163 int node = numa_cpu_lookup_table[cpu];
165 dbg("removing cpu %lu from node %d\n", cpu, node);
167 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
168 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
170 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
174 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
176 /* must hold reference to node during call */
177 static const __be32 *of_get_associativity(struct device_node *dev)
179 return of_get_property(dev, "ibm,associativity", NULL);
183 * Returns the property linux,drconf-usable-memory if
184 * it exists (the property exists only in kexec/kdump kernels,
185 * added by kexec-tools)
187 static const __be32 *of_get_usable_memory(struct device_node *memory)
191 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
192 if (!prop || len < sizeof(unsigned int))
197 int __node_distance(int a, int b)
200 int distance = LOCAL_DISTANCE;
203 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
205 for (i = 0; i < distance_ref_points_depth; i++) {
206 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
209 /* Double the distance for each NUMA level */
215 EXPORT_SYMBOL(__node_distance);
217 static void initialize_distance_lookup_table(int nid,
218 const __be32 *associativity)
225 for (i = 0; i < distance_ref_points_depth; i++) {
228 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
229 distance_lookup_table[nid][i] = of_read_number(entry, 1);
233 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
236 static int associativity_to_nid(const __be32 *associativity)
240 if (min_common_depth == -1)
243 if (of_read_number(associativity, 1) >= min_common_depth)
244 nid = of_read_number(&associativity[min_common_depth], 1);
246 /* POWER4 LPAR uses 0xffff as invalid node */
247 if (nid == 0xffff || nid >= MAX_NUMNODES)
251 of_read_number(associativity, 1) >= distance_ref_points_depth) {
253 * Skip the length field and send start of associativity array
255 initialize_distance_lookup_table(nid, associativity + 1);
262 /* Returns the nid associated with the given device tree node,
263 * or -1 if not found.
265 static int of_node_to_nid_single(struct device_node *device)
270 tmp = of_get_associativity(device);
272 nid = associativity_to_nid(tmp);
276 /* Walk the device tree upwards, looking for an associativity id */
277 int of_node_to_nid(struct device_node *device)
283 nid = of_node_to_nid_single(device);
287 device = of_get_next_parent(device);
293 EXPORT_SYMBOL(of_node_to_nid);
295 static int __init find_min_common_depth(void)
298 struct device_node *root;
300 if (firmware_has_feature(FW_FEATURE_OPAL))
301 root = of_find_node_by_path("/ibm,opal");
303 root = of_find_node_by_path("/rtas");
305 root = of_find_node_by_path("/");
308 * This property is a set of 32-bit integers, each representing
309 * an index into the ibm,associativity nodes.
311 * With form 0 affinity the first integer is for an SMP configuration
312 * (should be all 0's) and the second is for a normal NUMA
313 * configuration. We have only one level of NUMA.
315 * With form 1 affinity the first integer is the most significant
316 * NUMA boundary and the following are progressively less significant
317 * boundaries. There can be more than one level of NUMA.
319 distance_ref_points = of_get_property(root,
320 "ibm,associativity-reference-points",
321 &distance_ref_points_depth);
323 if (!distance_ref_points) {
324 dbg("NUMA: ibm,associativity-reference-points not found.\n");
328 distance_ref_points_depth /= sizeof(int);
330 if (firmware_has_feature(FW_FEATURE_OPAL) ||
331 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
332 dbg("Using form 1 affinity\n");
336 if (form1_affinity) {
337 depth = of_read_number(distance_ref_points, 1);
339 if (distance_ref_points_depth < 2) {
340 printk(KERN_WARNING "NUMA: "
341 "short ibm,associativity-reference-points\n");
345 depth = of_read_number(&distance_ref_points[1], 1);
349 * Warn and cap if the hardware supports more than
350 * MAX_DISTANCE_REF_POINTS domains.
352 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
353 printk(KERN_WARNING "NUMA: distance array capped at "
354 "%d entries\n", MAX_DISTANCE_REF_POINTS);
355 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
366 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
368 struct device_node *memory = NULL;
370 memory = of_find_node_by_type(memory, "memory");
372 panic("numa.c: No memory nodes found!");
374 *n_addr_cells = of_n_addr_cells(memory);
375 *n_size_cells = of_n_size_cells(memory);
379 static unsigned long read_n_cells(int n, const __be32 **buf)
381 unsigned long result = 0;
384 result = (result << 32) | of_read_number(*buf, 1);
391 * Read the next memblock list entry from the ibm,dynamic-memory property
392 * and return the information in the provided of_drconf_cell structure.
394 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
398 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
401 drmem->drc_index = of_read_number(cp, 1);
402 drmem->reserved = of_read_number(&cp[1], 1);
403 drmem->aa_index = of_read_number(&cp[2], 1);
404 drmem->flags = of_read_number(&cp[3], 1);
410 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
412 * The layout of the ibm,dynamic-memory property is a number N of memblock
413 * list entries followed by N memblock list entries. Each memblock list entry
414 * contains information as laid out in the of_drconf_cell struct above.
416 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
421 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
422 if (!prop || len < sizeof(unsigned int))
425 entries = of_read_number(prop++, 1);
427 /* Now that we know the number of entries, revalidate the size
428 * of the property read in to ensure we have everything
430 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
438 * Retrieve and validate the ibm,lmb-size property for drconf memory
439 * from the device tree.
441 static u64 of_get_lmb_size(struct device_node *memory)
446 prop = of_get_property(memory, "ibm,lmb-size", &len);
447 if (!prop || len < sizeof(unsigned int))
450 return read_n_cells(n_mem_size_cells, &prop);
453 struct assoc_arrays {
456 const __be32 *arrays;
460 * Retrieve and validate the list of associativity arrays for drconf
461 * memory from the ibm,associativity-lookup-arrays property of the
464 * The layout of the ibm,associativity-lookup-arrays property is a number N
465 * indicating the number of associativity arrays, followed by a number M
466 * indicating the size of each associativity array, followed by a list
467 * of N associativity arrays.
469 static int of_get_assoc_arrays(struct device_node *memory,
470 struct assoc_arrays *aa)
475 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
476 if (!prop || len < 2 * sizeof(unsigned int))
479 aa->n_arrays = of_read_number(prop++, 1);
480 aa->array_sz = of_read_number(prop++, 1);
482 /* Now that we know the number of arrays and size of each array,
483 * revalidate the size of the property read in.
485 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
493 * This is like of_node_to_nid_single() for memory represented in the
494 * ibm,dynamic-reconfiguration-memory node.
496 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
497 struct assoc_arrays *aa)
500 int nid = default_nid;
503 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
504 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
505 drmem->aa_index < aa->n_arrays) {
506 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
507 nid = of_read_number(&aa->arrays[index], 1);
509 if (nid == 0xffff || nid >= MAX_NUMNODES)
513 index = drmem->aa_index * aa->array_sz;
514 initialize_distance_lookup_table(nid,
523 * Figure out to which domain a cpu belongs and stick it there.
524 * Return the id of the domain used.
526 static int numa_setup_cpu(unsigned long lcpu)
529 struct device_node *cpu;
532 * If a valid cpu-to-node mapping is already available, use it
533 * directly instead of querying the firmware, since it represents
534 * the most recent mapping notified to us by the platform (eg: VPHN).
536 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
537 map_cpu_to_node(lcpu, nid);
541 cpu = of_get_cpu_node(lcpu, NULL);
545 if (cpu_present(lcpu))
551 nid = of_node_to_nid_single(cpu);
554 if (nid < 0 || !node_online(nid))
555 nid = first_online_node;
557 map_cpu_to_node(lcpu, nid);
563 static void verify_cpu_node_mapping(int cpu, int node)
565 int base, sibling, i;
567 /* Verify that all the threads in the core belong to the same node */
568 base = cpu_first_thread_sibling(cpu);
570 for (i = 0; i < threads_per_core; i++) {
573 if (sibling == cpu || cpu_is_offline(sibling))
576 if (cpu_to_node(sibling) != node) {
577 WARN(1, "CPU thread siblings %d and %d don't belong"
578 " to the same node!\n", cpu, sibling);
584 /* Must run before sched domains notifier. */
585 static int ppc_numa_cpu_prepare(unsigned int cpu)
589 nid = numa_setup_cpu(cpu);
590 verify_cpu_node_mapping(cpu, nid);
594 static int ppc_numa_cpu_dead(unsigned int cpu)
596 #ifdef CONFIG_HOTPLUG_CPU
597 unmap_cpu_from_node(cpu);
603 * Check and possibly modify a memory region to enforce the memory limit.
605 * Returns the size the region should have to enforce the memory limit.
606 * This will either be the original value of size, a truncated value,
607 * or zero. If the returned value of size is 0 the region should be
608 * discarded as it lies wholly above the memory limit.
610 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
614 * We use memblock_end_of_DRAM() in here instead of memory_limit because
615 * we've already adjusted it for the limit and it takes care of
616 * having memory holes below the limit. Also, in the case of
617 * iommu_is_off, memory_limit is not set but is implicitly enforced.
620 if (start + size <= memblock_end_of_DRAM())
623 if (start >= memblock_end_of_DRAM())
626 return memblock_end_of_DRAM() - start;
630 * Reads the counter for a given entry in
631 * linux,drconf-usable-memory property
633 static inline int __init read_usm_ranges(const __be32 **usm)
636 * For each lmb in ibm,dynamic-memory a corresponding
637 * entry in linux,drconf-usable-memory property contains
638 * a counter followed by that many (base, size) duple.
639 * read the counter from linux,drconf-usable-memory
641 return read_n_cells(n_mem_size_cells, usm);
645 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
646 * node. This assumes n_mem_{addr,size}_cells have been set.
648 static void __init parse_drconf_memory(struct device_node *memory)
650 const __be32 *uninitialized_var(dm), *usm;
651 unsigned int n, rc, ranges, is_kexec_kdump = 0;
652 unsigned long lmb_size, base, size, sz;
654 struct assoc_arrays aa = { .arrays = NULL };
656 n = of_get_drconf_memory(memory, &dm);
660 lmb_size = of_get_lmb_size(memory);
664 rc = of_get_assoc_arrays(memory, &aa);
668 /* check if this is a kexec/kdump kernel */
669 usm = of_get_usable_memory(memory);
673 for (; n != 0; --n) {
674 struct of_drconf_cell drmem;
676 read_drconf_cell(&drmem, &dm);
678 /* skip this block if the reserved bit is set in flags (0x80)
679 or if the block is not assigned to this partition (0x8) */
680 if ((drmem.flags & DRCONF_MEM_RESERVED)
681 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
684 base = drmem.base_addr;
688 if (is_kexec_kdump) {
689 ranges = read_usm_ranges(&usm);
690 if (!ranges) /* there are no (base, size) duple */
694 if (is_kexec_kdump) {
695 base = read_n_cells(n_mem_addr_cells, &usm);
696 size = read_n_cells(n_mem_size_cells, &usm);
698 nid = of_drconf_to_nid_single(&drmem, &aa);
699 fake_numa_create_new_node(
700 ((base + size) >> PAGE_SHIFT),
702 node_set_online(nid);
703 sz = numa_enforce_memory_limit(base, size);
705 memblock_set_node(base, sz,
706 &memblock.memory, nid);
711 static int __init parse_numa_properties(void)
713 struct device_node *memory;
717 if (numa_enabled == 0) {
718 printk(KERN_WARNING "NUMA disabled by user\n");
722 min_common_depth = find_min_common_depth();
724 if (min_common_depth < 0)
725 return min_common_depth;
727 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
730 * Even though we connect cpus to numa domains later in SMP
731 * init, we need to know the node ids now. This is because
732 * each node to be onlined must have NODE_DATA etc backing it.
734 for_each_present_cpu(i) {
735 struct device_node *cpu;
738 cpu = of_get_cpu_node(i, NULL);
740 nid = of_node_to_nid_single(cpu);
744 * Don't fall back to default_nid yet -- we will plug
745 * cpus into nodes once the memory scan has discovered
750 node_set_online(nid);
753 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
755 for_each_node_by_type(memory, "memory") {
760 const __be32 *memcell_buf;
763 memcell_buf = of_get_property(memory,
764 "linux,usable-memory", &len);
765 if (!memcell_buf || len <= 0)
766 memcell_buf = of_get_property(memory, "reg", &len);
767 if (!memcell_buf || len <= 0)
771 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
773 /* these are order-sensitive, and modify the buffer pointer */
774 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
775 size = read_n_cells(n_mem_size_cells, &memcell_buf);
778 * Assumption: either all memory nodes or none will
779 * have associativity properties. If none, then
780 * everything goes to default_nid.
782 nid = of_node_to_nid_single(memory);
786 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
787 node_set_online(nid);
789 size = numa_enforce_memory_limit(start, size);
791 memblock_set_node(start, size, &memblock.memory, nid);
798 * Now do the same thing for each MEMBLOCK listed in the
799 * ibm,dynamic-memory property in the
800 * ibm,dynamic-reconfiguration-memory node.
802 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
804 parse_drconf_memory(memory);
809 static void __init setup_nonnuma(void)
811 unsigned long top_of_ram = memblock_end_of_DRAM();
812 unsigned long total_ram = memblock_phys_mem_size();
813 unsigned long start_pfn, end_pfn;
814 unsigned int nid = 0;
815 struct memblock_region *reg;
817 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
818 top_of_ram, total_ram);
819 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
820 (top_of_ram - total_ram) >> 20);
822 for_each_memblock(memory, reg) {
823 start_pfn = memblock_region_memory_base_pfn(reg);
824 end_pfn = memblock_region_memory_end_pfn(reg);
826 fake_numa_create_new_node(end_pfn, &nid);
827 memblock_set_node(PFN_PHYS(start_pfn),
828 PFN_PHYS(end_pfn - start_pfn),
829 &memblock.memory, nid);
830 node_set_online(nid);
834 void __init dump_numa_cpu_topology(void)
837 unsigned int cpu, count;
839 if (min_common_depth == -1 || !numa_enabled)
842 for_each_online_node(node) {
843 pr_info("Node %d CPUs:", node);
847 * If we used a CPU iterator here we would miss printing
848 * the holes in the cpumap.
850 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
851 if (cpumask_test_cpu(cpu,
852 node_to_cpumask_map[node])) {
858 pr_cont("-%u", cpu - 1);
864 pr_cont("-%u", nr_cpu_ids - 1);
869 /* Initialize NODE_DATA for a node on the local memory */
870 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
872 u64 spanned_pages = end_pfn - start_pfn;
873 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
878 nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
881 /* report and initialize */
882 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
883 nd_pa, nd_pa + nd_size - 1);
884 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
886 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
889 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
890 NODE_DATA(nid)->node_id = nid;
891 NODE_DATA(nid)->node_start_pfn = start_pfn;
892 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
895 void __init initmem_init(void)
899 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
900 max_pfn = max_low_pfn;
902 if (parse_numa_properties())
908 * Reduce the possible NUMA nodes to the online NUMA nodes,
909 * since we do not support node hotplug. This ensures that we
910 * lower the maximum NUMA node ID to what is actually present.
912 nodes_and(node_possible_map, node_possible_map, node_online_map);
914 for_each_online_node(nid) {
915 unsigned long start_pfn, end_pfn;
917 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
918 setup_node_data(nid, start_pfn, end_pfn);
919 sparse_memory_present_with_active_regions(nid);
924 setup_node_to_cpumask_map();
926 reset_numa_cpu_lookup_table();
929 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
930 * even before we online them, so that we can use cpu_to_{node,mem}
931 * early in boot, cf. smp_prepare_cpus().
932 * _nocalls() + manual invocation is used because cpuhp is not yet
933 * initialized for the boot CPU.
935 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
936 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
937 for_each_present_cpu(cpu)
941 static int __init early_numa(char *p)
946 if (strstr(p, "off"))
949 if (strstr(p, "debug"))
952 p = strstr(p, "fake=");
954 cmdline = p + strlen("fake=");
958 early_param("numa", early_numa);
960 static bool topology_updates_enabled = true;
962 static int __init early_topology_updates(char *p)
967 if (!strcmp(p, "off")) {
968 pr_info("Disabling topology updates\n");
969 topology_updates_enabled = false;
974 early_param("topology_updates", early_topology_updates);
976 #ifdef CONFIG_MEMORY_HOTPLUG
978 * Find the node associated with a hot added memory section for
979 * memory represented in the device tree by the property
980 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
982 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
983 unsigned long scn_addr)
986 unsigned int drconf_cell_cnt, rc;
987 unsigned long lmb_size;
988 struct assoc_arrays aa;
991 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
992 if (!drconf_cell_cnt)
995 lmb_size = of_get_lmb_size(memory);
999 rc = of_get_assoc_arrays(memory, &aa);
1003 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1004 struct of_drconf_cell drmem;
1006 read_drconf_cell(&drmem, &dm);
1008 /* skip this block if it is reserved or not assigned to
1010 if ((drmem.flags & DRCONF_MEM_RESERVED)
1011 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1014 if ((scn_addr < drmem.base_addr)
1015 || (scn_addr >= (drmem.base_addr + lmb_size)))
1018 nid = of_drconf_to_nid_single(&drmem, &aa);
1026 * Find the node associated with a hot added memory section for memory
1027 * represented in the device tree as a node (i.e. memory@XXXX) for
1030 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1032 struct device_node *memory;
1035 for_each_node_by_type(memory, "memory") {
1036 unsigned long start, size;
1038 const __be32 *memcell_buf;
1041 memcell_buf = of_get_property(memory, "reg", &len);
1042 if (!memcell_buf || len <= 0)
1045 /* ranges in cell */
1046 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1049 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1050 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1052 if ((scn_addr < start) || (scn_addr >= (start + size)))
1055 nid = of_node_to_nid_single(memory);
1063 of_node_put(memory);
1069 * Find the node associated with a hot added memory section. Section
1070 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1071 * sections are fully contained within a single MEMBLOCK.
1073 int hot_add_scn_to_nid(unsigned long scn_addr)
1075 struct device_node *memory = NULL;
1078 if (!numa_enabled || (min_common_depth < 0))
1079 return first_online_node;
1081 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1083 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1084 of_node_put(memory);
1086 nid = hot_add_node_scn_to_nid(scn_addr);
1089 if (nid < 0 || !node_possible(nid))
1090 nid = first_online_node;
1095 static u64 hot_add_drconf_memory_max(void)
1097 struct device_node *memory = NULL;
1098 struct device_node *dn = NULL;
1099 unsigned int drconf_cell_cnt = 0;
1101 const __be32 *dm = NULL;
1102 const __be64 *lrdr = NULL;
1103 struct of_drconf_cell drmem;
1105 dn = of_find_node_by_path("/rtas");
1107 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1110 return be64_to_cpup(lrdr);
1113 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1115 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1116 lmb_size = of_get_lmb_size(memory);
1118 /* Advance to the last cell, each cell has 6 32 bit integers */
1119 dm += (drconf_cell_cnt - 1) * 6;
1120 read_drconf_cell(&drmem, &dm);
1121 of_node_put(memory);
1122 return drmem.base_addr + lmb_size;
1128 * memory_hotplug_max - return max address of memory that may be added
1130 * This is currently only used on systems that support drconfig memory
1133 u64 memory_hotplug_max(void)
1135 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1137 #endif /* CONFIG_MEMORY_HOTPLUG */
1139 /* Virtual Processor Home Node (VPHN) support */
1140 #ifdef CONFIG_PPC_SPLPAR
1144 struct topology_update_data {
1145 struct topology_update_data *next;
1151 #define TOPOLOGY_DEF_TIMER_SECS 60
1153 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1154 static cpumask_t cpu_associativity_changes_mask;
1155 static int vphn_enabled;
1156 static int prrn_enabled;
1157 static void reset_topology_timer(void);
1158 static int topology_timer_secs = 1;
1159 static int topology_inited;
1160 static int topology_update_needed;
1163 * Change polling interval for associativity changes.
1165 int timed_topology_update(int nsecs)
1169 topology_timer_secs = nsecs;
1171 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1173 reset_topology_timer();
1180 * Store the current values of the associativity change counters in the
1183 static void setup_cpu_associativity_change_counters(void)
1187 /* The VPHN feature supports a maximum of 8 reference points */
1188 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1190 for_each_possible_cpu(cpu) {
1192 u8 *counts = vphn_cpu_change_counts[cpu];
1193 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1195 for (i = 0; i < distance_ref_points_depth; i++)
1196 counts[i] = hypervisor_counts[i];
1201 * The hypervisor maintains a set of 8 associativity change counters in
1202 * the VPA of each cpu that correspond to the associativity levels in the
1203 * ibm,associativity-reference-points property. When an associativity
1204 * level changes, the corresponding counter is incremented.
1206 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1207 * node associativity levels have changed.
1209 * Returns the number of cpus with unhandled associativity changes.
1211 static int update_cpu_associativity_changes_mask(void)
1214 cpumask_t *changes = &cpu_associativity_changes_mask;
1216 for_each_possible_cpu(cpu) {
1218 u8 *counts = vphn_cpu_change_counts[cpu];
1219 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1221 for (i = 0; i < distance_ref_points_depth; i++) {
1222 if (hypervisor_counts[i] != counts[i]) {
1223 counts[i] = hypervisor_counts[i];
1228 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1229 cpu = cpu_last_thread_sibling(cpu);
1233 return cpumask_weight(changes);
1237 * Retrieve the new associativity information for a virtual processor's
1240 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1243 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1245 int hwcpu = get_hard_smp_processor_id(cpu);
1247 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1248 vphn_unpack_associativity(retbuf, associativity);
1253 static long vphn_get_associativity(unsigned long cpu,
1254 __be32 *associativity)
1258 rc = hcall_vphn(cpu, associativity);
1263 "VPHN is not supported. Disabling polling...\n");
1264 stop_topology_update();
1268 "hcall_vphn() experienced a hardware fault "
1269 "preventing VPHN. Disabling polling...\n");
1270 stop_topology_update();
1273 dbg("VPHN hcall succeeded. Reset polling...\n");
1274 timed_topology_update(0);
1282 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1283 * characteristics change. This function doesn't perform any locking and is
1284 * only safe to call from stop_machine().
1286 static int update_cpu_topology(void *data)
1288 struct topology_update_data *update;
1294 cpu = smp_processor_id();
1296 for (update = data; update; update = update->next) {
1297 int new_nid = update->new_nid;
1298 if (cpu != update->cpu)
1301 unmap_cpu_from_node(cpu);
1302 map_cpu_to_node(cpu, new_nid);
1303 set_cpu_numa_node(cpu, new_nid);
1304 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1311 static int update_lookup_table(void *data)
1313 struct topology_update_data *update;
1319 * Upon topology update, the numa-cpu lookup table needs to be updated
1320 * for all threads in the core, including offline CPUs, to ensure that
1321 * future hotplug operations respect the cpu-to-node associativity
1324 for (update = data; update; update = update->next) {
1327 nid = update->new_nid;
1328 base = cpu_first_thread_sibling(update->cpu);
1330 for (j = 0; j < threads_per_core; j++) {
1331 update_numa_cpu_lookup_table(base + j, nid);
1339 * Update the node maps and sysfs entries for each cpu whose home node
1340 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1342 * cpus_locked says whether we already hold cpu_hotplug_lock.
1344 int numa_update_cpu_topology(bool cpus_locked)
1346 unsigned int cpu, sibling, changed = 0;
1347 struct topology_update_data *updates, *ud;
1348 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1349 cpumask_t updated_cpus;
1351 int weight, new_nid, i = 0;
1353 if (!prrn_enabled && !vphn_enabled) {
1354 if (!topology_inited)
1355 topology_update_needed = 1;
1359 weight = cpumask_weight(&cpu_associativity_changes_mask);
1363 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1367 cpumask_clear(&updated_cpus);
1369 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1371 * If siblings aren't flagged for changes, updates list
1372 * will be too short. Skip on this update and set for next
1375 if (!cpumask_subset(cpu_sibling_mask(cpu),
1376 &cpu_associativity_changes_mask)) {
1377 pr_info("Sibling bits not set for associativity "
1378 "change, cpu%d\n", cpu);
1379 cpumask_or(&cpu_associativity_changes_mask,
1380 &cpu_associativity_changes_mask,
1381 cpu_sibling_mask(cpu));
1382 cpu = cpu_last_thread_sibling(cpu);
1386 /* Use associativity from first thread for all siblings */
1387 vphn_get_associativity(cpu, associativity);
1388 new_nid = associativity_to_nid(associativity);
1389 if (new_nid < 0 || !node_online(new_nid))
1390 new_nid = first_online_node;
1392 if (new_nid == numa_cpu_lookup_table[cpu]) {
1393 cpumask_andnot(&cpu_associativity_changes_mask,
1394 &cpu_associativity_changes_mask,
1395 cpu_sibling_mask(cpu));
1396 dbg("Assoc chg gives same node %d for cpu%d\n",
1398 cpu = cpu_last_thread_sibling(cpu);
1402 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1404 ud->next = &updates[i];
1406 ud->new_nid = new_nid;
1407 ud->old_nid = numa_cpu_lookup_table[sibling];
1408 cpumask_set_cpu(sibling, &updated_cpus);
1410 cpu = cpu_last_thread_sibling(cpu);
1414 * Prevent processing of 'updates' from overflowing array
1415 * where last entry filled in a 'next' pointer.
1418 updates[i-1].next = NULL;
1420 pr_debug("Topology update for the following CPUs:\n");
1421 if (cpumask_weight(&updated_cpus)) {
1422 for (ud = &updates[0]; ud; ud = ud->next) {
1423 pr_debug("cpu %d moving from node %d "
1425 ud->old_nid, ud->new_nid);
1430 * In cases where we have nothing to update (because the updates list
1431 * is too short or because the new topology is same as the old one),
1432 * skip invoking update_cpu_topology() via stop-machine(). This is
1433 * necessary (and not just a fast-path optimization) since stop-machine
1434 * can end up electing a random CPU to run update_cpu_topology(), and
1435 * thus trick us into setting up incorrect cpu-node mappings (since
1436 * 'updates' is kzalloc()'ed).
1438 * And for the similar reason, we will skip all the following updating.
1440 if (!cpumask_weight(&updated_cpus))
1444 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1447 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1450 * Update the numa-cpu lookup table with the new mappings, even for
1451 * offline CPUs. It is best to perform this update from the stop-
1455 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1456 cpumask_of(raw_smp_processor_id()));
1458 stop_machine(update_lookup_table, &updates[0],
1459 cpumask_of(raw_smp_processor_id()));
1461 for (ud = &updates[0]; ud; ud = ud->next) {
1462 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1463 register_cpu_under_node(ud->cpu, ud->new_nid);
1465 dev = get_cpu_device(ud->cpu);
1467 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1468 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1474 topology_update_needed = 0;
1478 int arch_update_cpu_topology(void)
1480 return numa_update_cpu_topology(true);
1483 static void topology_work_fn(struct work_struct *work)
1485 rebuild_sched_domains();
1487 static DECLARE_WORK(topology_work, topology_work_fn);
1489 static void topology_schedule_update(void)
1491 schedule_work(&topology_work);
1494 static void topology_timer_fn(struct timer_list *unused)
1496 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1497 topology_schedule_update();
1498 else if (vphn_enabled) {
1499 if (update_cpu_associativity_changes_mask() > 0)
1500 topology_schedule_update();
1501 reset_topology_timer();
1504 static struct timer_list topology_timer;
1506 static void reset_topology_timer(void)
1508 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1513 static void stage_topology_update(int core_id)
1515 cpumask_or(&cpu_associativity_changes_mask,
1516 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1517 reset_topology_timer();
1520 static int dt_update_callback(struct notifier_block *nb,
1521 unsigned long action, void *data)
1523 struct of_reconfig_data *update = data;
1524 int rc = NOTIFY_DONE;
1527 case OF_RECONFIG_UPDATE_PROPERTY:
1528 if (!of_prop_cmp(update->dn->type, "cpu") &&
1529 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1531 of_property_read_u32(update->dn, "reg", &core_id);
1532 stage_topology_update(core_id);
1541 static struct notifier_block dt_update_nb = {
1542 .notifier_call = dt_update_callback,
1548 * Start polling for associativity changes.
1550 int start_topology_update(void)
1554 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1555 if (!prrn_enabled) {
1558 rc = of_reconfig_notifier_register(&dt_update_nb);
1562 if (firmware_has_feature(FW_FEATURE_VPHN) &&
1563 lppaca_shared_proc(get_lppaca())) {
1564 if (!vphn_enabled) {
1566 setup_cpu_associativity_change_counters();
1567 timer_setup(&topology_timer, topology_timer_fn,
1569 reset_topology_timer();
1577 * Disable polling for VPHN associativity changes.
1579 int stop_topology_update(void)
1586 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1591 rc = del_timer_sync(&topology_timer);
1597 int prrn_is_enabled(void)
1599 return prrn_enabled;
1602 static int topology_read(struct seq_file *file, void *v)
1604 if (vphn_enabled || prrn_enabled)
1605 seq_puts(file, "on\n");
1607 seq_puts(file, "off\n");
1612 static int topology_open(struct inode *inode, struct file *file)
1614 return single_open(file, topology_read, NULL);
1617 static ssize_t topology_write(struct file *file, const char __user *buf,
1618 size_t count, loff_t *off)
1620 char kbuf[4]; /* "on" or "off" plus null. */
1623 read_len = count < 3 ? count : 3;
1624 if (copy_from_user(kbuf, buf, read_len))
1627 kbuf[read_len] = '\0';
1629 if (!strncmp(kbuf, "on", 2))
1630 start_topology_update();
1631 else if (!strncmp(kbuf, "off", 3))
1632 stop_topology_update();
1639 static const struct file_operations topology_ops = {
1641 .write = topology_write,
1642 .open = topology_open,
1643 .release = single_release
1646 static int topology_update_init(void)
1648 /* Do not poll for changes if disabled at boot */
1649 if (topology_updates_enabled)
1650 start_topology_update();
1653 topology_schedule_update();
1655 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1658 topology_inited = 1;
1659 if (topology_update_needed)
1660 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1665 device_initcall(topology_update_init);
1666 #endif /* CONFIG_PPC_SPLPAR */