4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
11 #define pr_fmt(fmt) "numa: " fmt
13 #include <linux/threads.h>
14 #include <linux/memblock.h>
15 #include <linux/init.h>
17 #include <linux/mmzone.h>
18 #include <linux/export.h>
19 #include <linux/nodemask.h>
20 #include <linux/cpu.h>
21 #include <linux/notifier.h>
23 #include <linux/pfn.h>
24 #include <linux/cpuset.h>
25 #include <linux/node.h>
26 #include <linux/stop_machine.h>
27 #include <linux/proc_fs.h>
28 #include <linux/seq_file.h>
29 #include <linux/uaccess.h>
30 #include <linux/slab.h>
31 #include <asm/cputhreads.h>
32 #include <asm/sparsemem.h>
35 #include <asm/topology.h>
36 #include <asm/firmware.h>
38 #include <asm/hvcall.h>
39 #include <asm/setup.h>
41 #include <asm/drmem.h>
43 static int numa_enabled = 1;
45 static char *cmdline __initdata;
47 static int numa_debug;
48 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
50 int numa_cpu_lookup_table[NR_CPUS];
51 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
52 struct pglist_data *node_data[MAX_NUMNODES];
54 EXPORT_SYMBOL(numa_cpu_lookup_table);
55 EXPORT_SYMBOL(node_to_cpumask_map);
56 EXPORT_SYMBOL(node_data);
58 static int min_common_depth;
59 static int n_mem_addr_cells, n_mem_size_cells;
60 static int form1_affinity;
62 #define MAX_DISTANCE_REF_POINTS 4
63 static int distance_ref_points_depth;
64 static const __be32 *distance_ref_points;
65 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
68 * Allocate node_to_cpumask_map based on number of available nodes
69 * Requires node_possible_map to be valid.
71 * Note: cpumask_of_node() is not valid until after this is done.
73 static void __init setup_node_to_cpumask_map(void)
77 /* setup nr_node_ids if not done yet */
78 if (nr_node_ids == MAX_NUMNODES)
81 /* allocate the map */
83 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
85 /* cpumask_of_node() will now work */
86 dbg("Node to cpumask map for %u nodes\n", nr_node_ids);
89 static int __init fake_numa_create_new_node(unsigned long end_pfn,
92 unsigned long long mem;
94 static unsigned int fake_nid;
95 static unsigned long long curr_boundary;
98 * Modify node id, iff we started creating NUMA nodes
99 * We want to continue from where we left of the last time
104 * In case there are no more arguments to parse, the
105 * node_id should be the same as the last fake node id
106 * (we've handled this above).
111 mem = memparse(p, &p);
115 if (mem < curr_boundary)
120 if ((end_pfn << PAGE_SHIFT) > mem) {
122 * Skip commas and spaces
124 while (*p == ',' || *p == ' ' || *p == '\t')
130 dbg("created new fake_node with id %d\n", fake_nid);
136 static void reset_numa_cpu_lookup_table(void)
140 for_each_possible_cpu(cpu)
141 numa_cpu_lookup_table[cpu] = -1;
144 static void map_cpu_to_node(int cpu, int node)
146 update_numa_cpu_lookup_table(cpu, node);
148 dbg("adding cpu %d to node %d\n", cpu, node);
150 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
151 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
154 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
155 static void unmap_cpu_from_node(unsigned long cpu)
157 int node = numa_cpu_lookup_table[cpu];
159 dbg("removing cpu %lu from node %d\n", cpu, node);
161 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
162 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
164 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
168 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
170 /* must hold reference to node during call */
171 static const __be32 *of_get_associativity(struct device_node *dev)
173 return of_get_property(dev, "ibm,associativity", NULL);
176 int __node_distance(int a, int b)
179 int distance = LOCAL_DISTANCE;
182 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
184 for (i = 0; i < distance_ref_points_depth; i++) {
185 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
188 /* Double the distance for each NUMA level */
194 EXPORT_SYMBOL(__node_distance);
196 static void initialize_distance_lookup_table(int nid,
197 const __be32 *associativity)
204 for (i = 0; i < distance_ref_points_depth; i++) {
207 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
208 distance_lookup_table[nid][i] = of_read_number(entry, 1);
212 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
215 static int associativity_to_nid(const __be32 *associativity)
217 int nid = NUMA_NO_NODE;
219 if (min_common_depth == -1)
222 if (of_read_number(associativity, 1) >= min_common_depth)
223 nid = of_read_number(&associativity[min_common_depth], 1);
225 /* POWER4 LPAR uses 0xffff as invalid node */
226 if (nid == 0xffff || nid >= MAX_NUMNODES)
230 of_read_number(associativity, 1) >= distance_ref_points_depth) {
232 * Skip the length field and send start of associativity array
234 initialize_distance_lookup_table(nid, associativity + 1);
241 /* Returns the nid associated with the given device tree node,
242 * or -1 if not found.
244 static int of_node_to_nid_single(struct device_node *device)
246 int nid = NUMA_NO_NODE;
249 tmp = of_get_associativity(device);
251 nid = associativity_to_nid(tmp);
255 /* Walk the device tree upwards, looking for an associativity id */
256 int of_node_to_nid(struct device_node *device)
258 int nid = NUMA_NO_NODE;
262 nid = of_node_to_nid_single(device);
266 device = of_get_next_parent(device);
272 EXPORT_SYMBOL(of_node_to_nid);
274 static int __init find_min_common_depth(void)
277 struct device_node *root;
279 if (firmware_has_feature(FW_FEATURE_OPAL))
280 root = of_find_node_by_path("/ibm,opal");
282 root = of_find_node_by_path("/rtas");
284 root = of_find_node_by_path("/");
287 * This property is a set of 32-bit integers, each representing
288 * an index into the ibm,associativity nodes.
290 * With form 0 affinity the first integer is for an SMP configuration
291 * (should be all 0's) and the second is for a normal NUMA
292 * configuration. We have only one level of NUMA.
294 * With form 1 affinity the first integer is the most significant
295 * NUMA boundary and the following are progressively less significant
296 * boundaries. There can be more than one level of NUMA.
298 distance_ref_points = of_get_property(root,
299 "ibm,associativity-reference-points",
300 &distance_ref_points_depth);
302 if (!distance_ref_points) {
303 dbg("NUMA: ibm,associativity-reference-points not found.\n");
307 distance_ref_points_depth /= sizeof(int);
309 if (firmware_has_feature(FW_FEATURE_OPAL) ||
310 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
311 dbg("Using form 1 affinity\n");
315 if (form1_affinity) {
316 depth = of_read_number(distance_ref_points, 1);
318 if (distance_ref_points_depth < 2) {
319 printk(KERN_WARNING "NUMA: "
320 "short ibm,associativity-reference-points\n");
324 depth = of_read_number(&distance_ref_points[1], 1);
328 * Warn and cap if the hardware supports more than
329 * MAX_DISTANCE_REF_POINTS domains.
331 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
332 printk(KERN_WARNING "NUMA: distance array capped at "
333 "%d entries\n", MAX_DISTANCE_REF_POINTS);
334 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
345 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
347 struct device_node *memory = NULL;
349 memory = of_find_node_by_type(memory, "memory");
351 panic("numa.c: No memory nodes found!");
353 *n_addr_cells = of_n_addr_cells(memory);
354 *n_size_cells = of_n_size_cells(memory);
358 static unsigned long read_n_cells(int n, const __be32 **buf)
360 unsigned long result = 0;
363 result = (result << 32) | of_read_number(*buf, 1);
369 struct assoc_arrays {
372 const __be32 *arrays;
376 * Retrieve and validate the list of associativity arrays for drconf
377 * memory from the ibm,associativity-lookup-arrays property of the
380 * The layout of the ibm,associativity-lookup-arrays property is a number N
381 * indicating the number of associativity arrays, followed by a number M
382 * indicating the size of each associativity array, followed by a list
383 * of N associativity arrays.
385 static int of_get_assoc_arrays(struct assoc_arrays *aa)
387 struct device_node *memory;
391 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
395 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
396 if (!prop || len < 2 * sizeof(unsigned int)) {
401 aa->n_arrays = of_read_number(prop++, 1);
402 aa->array_sz = of_read_number(prop++, 1);
406 /* Now that we know the number of arrays and size of each array,
407 * revalidate the size of the property read in.
409 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
417 * This is like of_node_to_nid_single() for memory represented in the
418 * ibm,dynamic-reconfiguration-memory node.
420 static int of_drconf_to_nid_single(struct drmem_lmb *lmb)
422 struct assoc_arrays aa = { .arrays = NULL };
424 int nid = default_nid;
427 rc = of_get_assoc_arrays(&aa);
431 if (min_common_depth > 0 && min_common_depth <= aa.array_sz &&
432 !(lmb->flags & DRCONF_MEM_AI_INVALID) &&
433 lmb->aa_index < aa.n_arrays) {
434 index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
435 nid = of_read_number(&aa.arrays[index], 1);
437 if (nid == 0xffff || nid >= MAX_NUMNODES)
441 index = lmb->aa_index * aa.array_sz;
442 initialize_distance_lookup_table(nid,
451 * Figure out to which domain a cpu belongs and stick it there.
452 * Return the id of the domain used.
454 static int numa_setup_cpu(unsigned long lcpu)
456 int nid = NUMA_NO_NODE;
457 struct device_node *cpu;
460 * If a valid cpu-to-node mapping is already available, use it
461 * directly instead of querying the firmware, since it represents
462 * the most recent mapping notified to us by the platform (eg: VPHN).
464 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
465 map_cpu_to_node(lcpu, nid);
469 cpu = of_get_cpu_node(lcpu, NULL);
473 if (cpu_present(lcpu))
479 nid = of_node_to_nid_single(cpu);
482 if (nid < 0 || !node_possible(nid))
483 nid = first_online_node;
485 map_cpu_to_node(lcpu, nid);
491 static void verify_cpu_node_mapping(int cpu, int node)
493 int base, sibling, i;
495 /* Verify that all the threads in the core belong to the same node */
496 base = cpu_first_thread_sibling(cpu);
498 for (i = 0; i < threads_per_core; i++) {
501 if (sibling == cpu || cpu_is_offline(sibling))
504 if (cpu_to_node(sibling) != node) {
505 WARN(1, "CPU thread siblings %d and %d don't belong"
506 " to the same node!\n", cpu, sibling);
512 /* Must run before sched domains notifier. */
513 static int ppc_numa_cpu_prepare(unsigned int cpu)
517 nid = numa_setup_cpu(cpu);
518 verify_cpu_node_mapping(cpu, nid);
522 static int ppc_numa_cpu_dead(unsigned int cpu)
524 #ifdef CONFIG_HOTPLUG_CPU
525 unmap_cpu_from_node(cpu);
531 * Check and possibly modify a memory region to enforce the memory limit.
533 * Returns the size the region should have to enforce the memory limit.
534 * This will either be the original value of size, a truncated value,
535 * or zero. If the returned value of size is 0 the region should be
536 * discarded as it lies wholly above the memory limit.
538 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
542 * We use memblock_end_of_DRAM() in here instead of memory_limit because
543 * we've already adjusted it for the limit and it takes care of
544 * having memory holes below the limit. Also, in the case of
545 * iommu_is_off, memory_limit is not set but is implicitly enforced.
548 if (start + size <= memblock_end_of_DRAM())
551 if (start >= memblock_end_of_DRAM())
554 return memblock_end_of_DRAM() - start;
558 * Reads the counter for a given entry in
559 * linux,drconf-usable-memory property
561 static inline int __init read_usm_ranges(const __be32 **usm)
564 * For each lmb in ibm,dynamic-memory a corresponding
565 * entry in linux,drconf-usable-memory property contains
566 * a counter followed by that many (base, size) duple.
567 * read the counter from linux,drconf-usable-memory
569 return read_n_cells(n_mem_size_cells, usm);
573 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
574 * node. This assumes n_mem_{addr,size}_cells have been set.
576 static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
579 unsigned int ranges, is_kexec_kdump = 0;
580 unsigned long base, size, sz;
584 * Skip this block if the reserved bit is set in flags (0x80)
585 * or if the block is not assigned to this partition (0x8)
587 if ((lmb->flags & DRCONF_MEM_RESERVED)
588 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
594 base = lmb->base_addr;
595 size = drmem_lmb_size();
598 if (is_kexec_kdump) {
599 ranges = read_usm_ranges(usm);
600 if (!ranges) /* there are no (base, size) duple */
605 if (is_kexec_kdump) {
606 base = read_n_cells(n_mem_addr_cells, usm);
607 size = read_n_cells(n_mem_size_cells, usm);
610 nid = of_drconf_to_nid_single(lmb);
611 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
613 node_set_online(nid);
614 sz = numa_enforce_memory_limit(base, size);
616 memblock_set_node(base, sz, &memblock.memory, nid);
620 static int __init parse_numa_properties(void)
622 struct device_node *memory;
626 if (numa_enabled == 0) {
627 printk(KERN_WARNING "NUMA disabled by user\n");
631 min_common_depth = find_min_common_depth();
633 if (min_common_depth < 0)
634 return min_common_depth;
636 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
639 * Even though we connect cpus to numa domains later in SMP
640 * init, we need to know the node ids now. This is because
641 * each node to be onlined must have NODE_DATA etc backing it.
643 for_each_present_cpu(i) {
644 struct device_node *cpu;
647 cpu = of_get_cpu_node(i, NULL);
649 nid = of_node_to_nid_single(cpu);
653 * Don't fall back to default_nid yet -- we will plug
654 * cpus into nodes once the memory scan has discovered
659 node_set_online(nid);
662 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
664 for_each_node_by_type(memory, "memory") {
669 const __be32 *memcell_buf;
672 memcell_buf = of_get_property(memory,
673 "linux,usable-memory", &len);
674 if (!memcell_buf || len <= 0)
675 memcell_buf = of_get_property(memory, "reg", &len);
676 if (!memcell_buf || len <= 0)
680 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
682 /* these are order-sensitive, and modify the buffer pointer */
683 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
684 size = read_n_cells(n_mem_size_cells, &memcell_buf);
687 * Assumption: either all memory nodes or none will
688 * have associativity properties. If none, then
689 * everything goes to default_nid.
691 nid = of_node_to_nid_single(memory);
695 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
696 node_set_online(nid);
698 size = numa_enforce_memory_limit(start, size);
700 memblock_set_node(start, size, &memblock.memory, nid);
707 * Now do the same thing for each MEMBLOCK listed in the
708 * ibm,dynamic-memory property in the
709 * ibm,dynamic-reconfiguration-memory node.
711 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
713 walk_drmem_lmbs(memory, numa_setup_drmem_lmb);
720 static void __init setup_nonnuma(void)
722 unsigned long top_of_ram = memblock_end_of_DRAM();
723 unsigned long total_ram = memblock_phys_mem_size();
724 unsigned long start_pfn, end_pfn;
725 unsigned int nid = 0;
726 struct memblock_region *reg;
728 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
729 top_of_ram, total_ram);
730 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
731 (top_of_ram - total_ram) >> 20);
733 for_each_memblock(memory, reg) {
734 start_pfn = memblock_region_memory_base_pfn(reg);
735 end_pfn = memblock_region_memory_end_pfn(reg);
737 fake_numa_create_new_node(end_pfn, &nid);
738 memblock_set_node(PFN_PHYS(start_pfn),
739 PFN_PHYS(end_pfn - start_pfn),
740 &memblock.memory, nid);
741 node_set_online(nid);
745 void __init dump_numa_cpu_topology(void)
748 unsigned int cpu, count;
750 if (min_common_depth == -1 || !numa_enabled)
753 for_each_online_node(node) {
754 pr_info("Node %d CPUs:", node);
758 * If we used a CPU iterator here we would miss printing
759 * the holes in the cpumap.
761 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
762 if (cpumask_test_cpu(cpu,
763 node_to_cpumask_map[node])) {
769 pr_cont("-%u", cpu - 1);
775 pr_cont("-%u", nr_cpu_ids - 1);
780 /* Initialize NODE_DATA for a node on the local memory */
781 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
783 u64 spanned_pages = end_pfn - start_pfn;
784 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
789 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
791 panic("Cannot allocate %zu bytes for node %d data\n",
796 /* report and initialize */
797 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
798 nd_pa, nd_pa + nd_size - 1);
799 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
801 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
804 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
805 NODE_DATA(nid)->node_id = nid;
806 NODE_DATA(nid)->node_start_pfn = start_pfn;
807 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
810 static void __init find_possible_nodes(void)
812 struct device_node *rtas;
815 if (min_common_depth <= 0)
818 rtas = of_find_node_by_path("/rtas");
822 if (of_property_read_u32_index(rtas,
823 "ibm,max-associativity-domains",
824 min_common_depth, &numnodes))
827 for (i = 0; i < numnodes; i++) {
828 if (!node_possible(i))
829 node_set(i, node_possible_map);
836 void __init mem_topology_setup(void)
840 if (parse_numa_properties())
844 * Modify the set of possible NUMA nodes to reflect information
845 * available about the set of online nodes, and the set of nodes
846 * that we expect to make use of for this platform's affinity
849 nodes_and(node_possible_map, node_possible_map, node_online_map);
851 find_possible_nodes();
853 setup_node_to_cpumask_map();
855 reset_numa_cpu_lookup_table();
857 for_each_present_cpu(cpu)
861 void __init initmem_init(void)
865 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
866 max_pfn = max_low_pfn;
870 for_each_online_node(nid) {
871 unsigned long start_pfn, end_pfn;
873 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
874 setup_node_data(nid, start_pfn, end_pfn);
875 sparse_memory_present_with_active_regions(nid);
881 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
882 * even before we online them, so that we can use cpu_to_{node,mem}
883 * early in boot, cf. smp_prepare_cpus().
884 * _nocalls() + manual invocation is used because cpuhp is not yet
885 * initialized for the boot CPU.
887 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
888 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
891 static int __init early_numa(char *p)
896 if (strstr(p, "off"))
899 if (strstr(p, "debug"))
902 p = strstr(p, "fake=");
904 cmdline = p + strlen("fake=");
908 early_param("numa", early_numa);
911 * The platform can inform us through one of several mechanisms
912 * (post-migration device tree updates, PRRN or VPHN) that the NUMA
913 * assignment of a resource has changed. This controls whether we act
914 * on that. Disabled by default.
916 static bool topology_updates_enabled;
918 static int __init early_topology_updates(char *p)
923 if (!strcmp(p, "on")) {
924 pr_warn("Caution: enabling topology updates\n");
925 topology_updates_enabled = true;
930 early_param("topology_updates", early_topology_updates);
932 #ifdef CONFIG_MEMORY_HOTPLUG
934 * Find the node associated with a hot added memory section for
935 * memory represented in the device tree by the property
936 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
938 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
940 struct drmem_lmb *lmb;
941 unsigned long lmb_size;
942 int nid = NUMA_NO_NODE;
944 lmb_size = drmem_lmb_size();
946 for_each_drmem_lmb(lmb) {
947 /* skip this block if it is reserved or not assigned to
949 if ((lmb->flags & DRCONF_MEM_RESERVED)
950 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
953 if ((scn_addr < lmb->base_addr)
954 || (scn_addr >= (lmb->base_addr + lmb_size)))
957 nid = of_drconf_to_nid_single(lmb);
965 * Find the node associated with a hot added memory section for memory
966 * represented in the device tree as a node (i.e. memory@XXXX) for
969 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
971 struct device_node *memory;
972 int nid = NUMA_NO_NODE;
974 for_each_node_by_type(memory, "memory") {
975 unsigned long start, size;
977 const __be32 *memcell_buf;
980 memcell_buf = of_get_property(memory, "reg", &len);
981 if (!memcell_buf || len <= 0)
985 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
988 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
989 size = read_n_cells(n_mem_size_cells, &memcell_buf);
991 if ((scn_addr < start) || (scn_addr >= (start + size)))
994 nid = of_node_to_nid_single(memory);
1002 of_node_put(memory);
1008 * Find the node associated with a hot added memory section. Section
1009 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1010 * sections are fully contained within a single MEMBLOCK.
1012 int hot_add_scn_to_nid(unsigned long scn_addr)
1014 struct device_node *memory = NULL;
1017 if (!numa_enabled || (min_common_depth < 0))
1018 return first_online_node;
1020 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1022 nid = hot_add_drconf_scn_to_nid(scn_addr);
1023 of_node_put(memory);
1025 nid = hot_add_node_scn_to_nid(scn_addr);
1028 if (nid < 0 || !node_possible(nid))
1029 nid = first_online_node;
1034 static u64 hot_add_drconf_memory_max(void)
1036 struct device_node *memory = NULL;
1037 struct device_node *dn = NULL;
1038 const __be64 *lrdr = NULL;
1040 dn = of_find_node_by_path("/rtas");
1042 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1045 return be64_to_cpup(lrdr);
1048 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1050 of_node_put(memory);
1051 return drmem_lmb_memory_max();
1057 * memory_hotplug_max - return max address of memory that may be added
1059 * This is currently only used on systems that support drconfig memory
1062 u64 memory_hotplug_max(void)
1064 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1066 #endif /* CONFIG_MEMORY_HOTPLUG */
1068 /* Virtual Processor Home Node (VPHN) support */
1069 #ifdef CONFIG_PPC_SPLPAR
1071 #include "book3s64/vphn.h"
1073 struct topology_update_data {
1074 struct topology_update_data *next;
1080 #define TOPOLOGY_DEF_TIMER_SECS 60
1082 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1083 static cpumask_t cpu_associativity_changes_mask;
1084 static int vphn_enabled;
1085 static int prrn_enabled;
1086 static void reset_topology_timer(void);
1087 static int topology_timer_secs = 1;
1088 static int topology_inited;
1091 * Change polling interval for associativity changes.
1093 int timed_topology_update(int nsecs)
1097 topology_timer_secs = nsecs;
1099 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1101 reset_topology_timer();
1108 * Store the current values of the associativity change counters in the
1111 static void setup_cpu_associativity_change_counters(void)
1115 /* The VPHN feature supports a maximum of 8 reference points */
1116 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1118 for_each_possible_cpu(cpu) {
1120 u8 *counts = vphn_cpu_change_counts[cpu];
1121 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1123 for (i = 0; i < distance_ref_points_depth; i++)
1124 counts[i] = hypervisor_counts[i];
1129 * The hypervisor maintains a set of 8 associativity change counters in
1130 * the VPA of each cpu that correspond to the associativity levels in the
1131 * ibm,associativity-reference-points property. When an associativity
1132 * level changes, the corresponding counter is incremented.
1134 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1135 * node associativity levels have changed.
1137 * Returns the number of cpus with unhandled associativity changes.
1139 static int update_cpu_associativity_changes_mask(void)
1142 cpumask_t *changes = &cpu_associativity_changes_mask;
1144 for_each_possible_cpu(cpu) {
1146 u8 *counts = vphn_cpu_change_counts[cpu];
1147 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1149 for (i = 0; i < distance_ref_points_depth; i++) {
1150 if (hypervisor_counts[i] != counts[i]) {
1151 counts[i] = hypervisor_counts[i];
1156 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1157 cpu = cpu_last_thread_sibling(cpu);
1161 return cpumask_weight(changes);
1165 * Retrieve the new associativity information for a virtual processor's
1168 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1171 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1173 int hwcpu = get_hard_smp_processor_id(cpu);
1175 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1176 vphn_unpack_associativity(retbuf, associativity);
1181 static long vphn_get_associativity(unsigned long cpu,
1182 __be32 *associativity)
1186 rc = hcall_vphn(cpu, associativity);
1190 printk_once(KERN_INFO
1191 "VPHN is not supported. Disabling polling...\n");
1192 stop_topology_update();
1196 "hcall_vphn() experienced a hardware fault "
1197 "preventing VPHN. Disabling polling...\n");
1198 stop_topology_update();
1201 dbg("VPHN hcall succeeded. Reset polling...\n");
1202 timed_topology_update(0);
1209 int find_and_online_cpu_nid(int cpu)
1211 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1214 /* Use associativity from first thread for all siblings */
1215 if (vphn_get_associativity(cpu, associativity))
1216 return cpu_to_node(cpu);
1218 new_nid = associativity_to_nid(associativity);
1219 if (new_nid < 0 || !node_possible(new_nid))
1220 new_nid = first_online_node;
1222 if (NODE_DATA(new_nid) == NULL) {
1223 #ifdef CONFIG_MEMORY_HOTPLUG
1225 * Need to ensure that NODE_DATA is initialized for a node from
1226 * available memory (see memblock_alloc_try_nid). If unable to
1227 * init the node, then default to nearest node that has memory
1228 * installed. Skip onlining a node if the subsystems are not
1231 if (!topology_inited || try_online_node(new_nid))
1232 new_nid = first_online_node;
1235 * Default to using the nearest node that has memory installed.
1236 * Otherwise, it would be necessary to patch the kernel MM code
1237 * to deal with more memoryless-node error conditions.
1239 new_nid = first_online_node;
1243 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1249 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1250 * characteristics change. This function doesn't perform any locking and is
1251 * only safe to call from stop_machine().
1253 static int update_cpu_topology(void *data)
1255 struct topology_update_data *update;
1261 cpu = smp_processor_id();
1263 for (update = data; update; update = update->next) {
1264 int new_nid = update->new_nid;
1265 if (cpu != update->cpu)
1268 unmap_cpu_from_node(cpu);
1269 map_cpu_to_node(cpu, new_nid);
1270 set_cpu_numa_node(cpu, new_nid);
1271 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1278 static int update_lookup_table(void *data)
1280 struct topology_update_data *update;
1286 * Upon topology update, the numa-cpu lookup table needs to be updated
1287 * for all threads in the core, including offline CPUs, to ensure that
1288 * future hotplug operations respect the cpu-to-node associativity
1291 for (update = data; update; update = update->next) {
1294 nid = update->new_nid;
1295 base = cpu_first_thread_sibling(update->cpu);
1297 for (j = 0; j < threads_per_core; j++) {
1298 update_numa_cpu_lookup_table(base + j, nid);
1306 * Update the node maps and sysfs entries for each cpu whose home node
1307 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1309 * cpus_locked says whether we already hold cpu_hotplug_lock.
1311 int numa_update_cpu_topology(bool cpus_locked)
1313 unsigned int cpu, sibling, changed = 0;
1314 struct topology_update_data *updates, *ud;
1315 cpumask_t updated_cpus;
1317 int weight, new_nid, i = 0;
1319 if (!prrn_enabled && !vphn_enabled && topology_inited)
1322 weight = cpumask_weight(&cpu_associativity_changes_mask);
1326 updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL);
1330 cpumask_clear(&updated_cpus);
1332 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1334 * If siblings aren't flagged for changes, updates list
1335 * will be too short. Skip on this update and set for next
1338 if (!cpumask_subset(cpu_sibling_mask(cpu),
1339 &cpu_associativity_changes_mask)) {
1340 pr_info("Sibling bits not set for associativity "
1341 "change, cpu%d\n", cpu);
1342 cpumask_or(&cpu_associativity_changes_mask,
1343 &cpu_associativity_changes_mask,
1344 cpu_sibling_mask(cpu));
1345 cpu = cpu_last_thread_sibling(cpu);
1349 new_nid = find_and_online_cpu_nid(cpu);
1351 if (new_nid == numa_cpu_lookup_table[cpu]) {
1352 cpumask_andnot(&cpu_associativity_changes_mask,
1353 &cpu_associativity_changes_mask,
1354 cpu_sibling_mask(cpu));
1355 dbg("Assoc chg gives same node %d for cpu%d\n",
1357 cpu = cpu_last_thread_sibling(cpu);
1361 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1363 ud->next = &updates[i];
1365 ud->new_nid = new_nid;
1366 ud->old_nid = numa_cpu_lookup_table[sibling];
1367 cpumask_set_cpu(sibling, &updated_cpus);
1369 cpu = cpu_last_thread_sibling(cpu);
1373 * Prevent processing of 'updates' from overflowing array
1374 * where last entry filled in a 'next' pointer.
1377 updates[i-1].next = NULL;
1379 pr_debug("Topology update for the following CPUs:\n");
1380 if (cpumask_weight(&updated_cpus)) {
1381 for (ud = &updates[0]; ud; ud = ud->next) {
1382 pr_debug("cpu %d moving from node %d "
1384 ud->old_nid, ud->new_nid);
1389 * In cases where we have nothing to update (because the updates list
1390 * is too short or because the new topology is same as the old one),
1391 * skip invoking update_cpu_topology() via stop-machine(). This is
1392 * necessary (and not just a fast-path optimization) since stop-machine
1393 * can end up electing a random CPU to run update_cpu_topology(), and
1394 * thus trick us into setting up incorrect cpu-node mappings (since
1395 * 'updates' is kzalloc()'ed).
1397 * And for the similar reason, we will skip all the following updating.
1399 if (!cpumask_weight(&updated_cpus))
1403 stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1406 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1409 * Update the numa-cpu lookup table with the new mappings, even for
1410 * offline CPUs. It is best to perform this update from the stop-
1414 stop_machine_cpuslocked(update_lookup_table, &updates[0],
1415 cpumask_of(raw_smp_processor_id()));
1417 stop_machine(update_lookup_table, &updates[0],
1418 cpumask_of(raw_smp_processor_id()));
1420 for (ud = &updates[0]; ud; ud = ud->next) {
1421 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1422 register_cpu_under_node(ud->cpu, ud->new_nid);
1424 dev = get_cpu_device(ud->cpu);
1426 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1427 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1436 int arch_update_cpu_topology(void)
1438 return numa_update_cpu_topology(true);
1441 static void topology_work_fn(struct work_struct *work)
1443 rebuild_sched_domains();
1445 static DECLARE_WORK(topology_work, topology_work_fn);
1447 static void topology_schedule_update(void)
1449 schedule_work(&topology_work);
1452 static void topology_timer_fn(struct timer_list *unused)
1454 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1455 topology_schedule_update();
1456 else if (vphn_enabled) {
1457 if (update_cpu_associativity_changes_mask() > 0)
1458 topology_schedule_update();
1459 reset_topology_timer();
1462 static struct timer_list topology_timer;
1464 static void reset_topology_timer(void)
1467 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1472 static int dt_update_callback(struct notifier_block *nb,
1473 unsigned long action, void *data)
1475 struct of_reconfig_data *update = data;
1476 int rc = NOTIFY_DONE;
1479 case OF_RECONFIG_UPDATE_PROPERTY:
1480 if (of_node_is_type(update->dn, "cpu") &&
1481 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1483 of_property_read_u32(update->dn, "reg", &core_id);
1484 rc = dlpar_cpu_readd(core_id);
1493 static struct notifier_block dt_update_nb = {
1494 .notifier_call = dt_update_callback,
1500 * Start polling for associativity changes.
1502 int start_topology_update(void)
1506 if (!topology_updates_enabled)
1509 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1510 if (!prrn_enabled) {
1513 rc = of_reconfig_notifier_register(&dt_update_nb);
1517 if (firmware_has_feature(FW_FEATURE_VPHN) &&
1518 lppaca_shared_proc(get_lppaca())) {
1519 if (!vphn_enabled) {
1521 setup_cpu_associativity_change_counters();
1522 timer_setup(&topology_timer, topology_timer_fn,
1524 reset_topology_timer();
1528 pr_info("Starting topology update%s%s\n",
1529 (prrn_enabled ? " prrn_enabled" : ""),
1530 (vphn_enabled ? " vphn_enabled" : ""));
1536 * Disable polling for VPHN associativity changes.
1538 int stop_topology_update(void)
1542 if (!topology_updates_enabled)
1548 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1553 rc = del_timer_sync(&topology_timer);
1556 pr_info("Stopping topology update\n");
1561 int prrn_is_enabled(void)
1563 return prrn_enabled;
1566 void __init shared_proc_topology_init(void)
1568 if (lppaca_shared_proc(get_lppaca())) {
1569 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1571 numa_update_cpu_topology(false);
1575 static int topology_read(struct seq_file *file, void *v)
1577 if (vphn_enabled || prrn_enabled)
1578 seq_puts(file, "on\n");
1580 seq_puts(file, "off\n");
1585 static int topology_open(struct inode *inode, struct file *file)
1587 return single_open(file, topology_read, NULL);
1590 static ssize_t topology_write(struct file *file, const char __user *buf,
1591 size_t count, loff_t *off)
1593 char kbuf[4]; /* "on" or "off" plus null. */
1596 read_len = count < 3 ? count : 3;
1597 if (copy_from_user(kbuf, buf, read_len))
1600 kbuf[read_len] = '\0';
1602 if (!strncmp(kbuf, "on", 2)) {
1603 topology_updates_enabled = true;
1604 start_topology_update();
1605 } else if (!strncmp(kbuf, "off", 3)) {
1606 stop_topology_update();
1607 topology_updates_enabled = false;
1614 static const struct file_operations topology_ops = {
1616 .write = topology_write,
1617 .open = topology_open,
1618 .release = single_release
1621 static int topology_update_init(void)
1623 start_topology_update();
1626 topology_schedule_update();
1628 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1631 topology_inited = 1;
1634 device_initcall(topology_update_init);
1635 #endif /* CONFIG_PPC_SPLPAR */