1 // SPDX-License-Identifier: GPL-2.0-or-later
5 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
7 #define pr_fmt(fmt) "numa: " fmt
9 #include <linux/threads.h>
10 #include <linux/memblock.h>
11 #include <linux/init.h>
13 #include <linux/mmzone.h>
14 #include <linux/export.h>
15 #include <linux/nodemask.h>
16 #include <linux/cpu.h>
17 #include <linux/notifier.h>
19 #include <linux/pfn.h>
20 #include <linux/cpuset.h>
21 #include <linux/node.h>
22 #include <linux/stop_machine.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/uaccess.h>
26 #include <linux/slab.h>
27 #include <asm/cputhreads.h>
28 #include <asm/sparsemem.h>
31 #include <asm/topology.h>
32 #include <asm/firmware.h>
34 #include <asm/hvcall.h>
35 #include <asm/setup.h>
37 #include <asm/drmem.h>
39 static int numa_enabled = 1;
41 static char *cmdline __initdata;
43 int numa_cpu_lookup_table[NR_CPUS];
44 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
45 struct pglist_data *node_data[MAX_NUMNODES];
47 EXPORT_SYMBOL(numa_cpu_lookup_table);
48 EXPORT_SYMBOL(node_to_cpumask_map);
49 EXPORT_SYMBOL(node_data);
51 static int primary_domain_index;
52 static int n_mem_addr_cells, n_mem_size_cells;
54 #define FORM0_AFFINITY 0
55 #define FORM1_AFFINITY 1
56 #define FORM2_AFFINITY 2
57 static int affinity_form;
59 #define MAX_DISTANCE_REF_POINTS 4
60 static int distance_ref_points_depth;
61 static const __be32 *distance_ref_points;
62 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
63 static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
64 [0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
66 static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
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 pr_debug("Node to cpumask map for %u 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 pr_debug("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 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) {
150 pr_debug("adding cpu %d to node %d\n", cpu, node);
151 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 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
161 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
162 pr_debug("removing cpu %lu from node %d\n", cpu, node);
164 pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node);
167 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
169 static int __associativity_to_nid(const __be32 *associativity,
174 * primary_domain_index is 1 based array index.
176 int index = primary_domain_index - 1;
178 if (!numa_enabled || index >= max_array_sz)
181 nid = of_read_number(&associativity[index], 1);
183 /* POWER4 LPAR uses 0xffff as invalid node */
184 if (nid == 0xffff || nid >= nr_node_ids)
189 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
192 static int associativity_to_nid(const __be32 *associativity)
194 int array_sz = of_read_number(associativity, 1);
196 /* Skip the first element in the associativity array */
197 return __associativity_to_nid((associativity + 1), array_sz);
200 static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
205 node1 = associativity_to_nid(cpu1_assoc);
206 node2 = associativity_to_nid(cpu2_assoc);
208 dist = numa_distance_table[node1][node2];
209 if (dist <= LOCAL_DISTANCE)
211 else if (dist <= REMOTE_DISTANCE)
217 static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
223 for (i = 0; i < distance_ref_points_depth; i++) {
224 index = be32_to_cpu(distance_ref_points[i]);
225 if (cpu1_assoc[index] == cpu2_assoc[index])
233 int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
235 /* We should not get called with FORM0 */
236 VM_WARN_ON(affinity_form == FORM0_AFFINITY);
237 if (affinity_form == FORM1_AFFINITY)
238 return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
239 return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
242 /* must hold reference to node during call */
243 static const __be32 *of_get_associativity(struct device_node *dev)
245 return of_get_property(dev, "ibm,associativity", NULL);
248 int __node_distance(int a, int b)
251 int distance = LOCAL_DISTANCE;
253 if (affinity_form == FORM2_AFFINITY)
254 return numa_distance_table[a][b];
255 else if (affinity_form == FORM0_AFFINITY)
256 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
258 for (i = 0; i < distance_ref_points_depth; i++) {
259 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
262 /* Double the distance for each NUMA level */
268 EXPORT_SYMBOL(__node_distance);
270 /* Returns the nid associated with the given device tree node,
271 * or -1 if not found.
273 static int of_node_to_nid_single(struct device_node *device)
275 int nid = NUMA_NO_NODE;
278 tmp = of_get_associativity(device);
280 nid = associativity_to_nid(tmp);
284 /* Walk the device tree upwards, looking for an associativity id */
285 int of_node_to_nid(struct device_node *device)
287 int nid = NUMA_NO_NODE;
291 nid = of_node_to_nid_single(device);
295 device = of_get_next_parent(device);
301 EXPORT_SYMBOL(of_node_to_nid);
303 static void __initialize_form1_numa_distance(const __be32 *associativity,
308 if (affinity_form != FORM1_AFFINITY)
311 nid = __associativity_to_nid(associativity, max_array_sz);
312 if (nid != NUMA_NO_NODE) {
313 for (i = 0; i < distance_ref_points_depth; i++) {
315 int index = be32_to_cpu(distance_ref_points[i]) - 1;
318 * broken hierarchy, return with broken distance table
320 if (WARN(index >= max_array_sz, "Broken ibm,associativity property"))
323 entry = &associativity[index];
324 distance_lookup_table[nid][i] = of_read_number(entry, 1);
329 static void initialize_form1_numa_distance(const __be32 *associativity)
333 array_sz = of_read_number(associativity, 1);
334 /* Skip the first element in the associativity array */
335 __initialize_form1_numa_distance(associativity + 1, array_sz);
339 * Used to update distance information w.r.t newly added node.
341 void update_numa_distance(struct device_node *node)
345 if (affinity_form == FORM0_AFFINITY)
347 else if (affinity_form == FORM1_AFFINITY) {
348 const __be32 *associativity;
350 associativity = of_get_associativity(node);
354 initialize_form1_numa_distance(associativity);
359 nid = of_node_to_nid_single(node);
360 if (nid == NUMA_NO_NODE)
364 * With FORM2 we expect NUMA distance of all possible NUMA
365 * nodes to be provided during boot.
367 WARN(numa_distance_table[nid][nid] == -1,
368 "NUMA distance details for node %d not provided\n", nid);
372 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
373 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
375 static void initialize_form2_numa_distance_lookup_table(void)
378 struct device_node *root;
379 const __u8 *numa_dist_table;
380 const __be32 *numa_lookup_index;
381 int numa_dist_table_length;
382 int max_numa_index, distance_index;
384 if (firmware_has_feature(FW_FEATURE_OPAL))
385 root = of_find_node_by_path("/ibm,opal");
387 root = of_find_node_by_path("/rtas");
389 root = of_find_node_by_path("/");
391 numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
392 max_numa_index = of_read_number(&numa_lookup_index[0], 1);
394 /* first element of the array is the size and is encode-int */
395 numa_dist_table = of_get_property(root, "ibm,numa-distance-table", NULL);
396 numa_dist_table_length = of_read_number((const __be32 *)&numa_dist_table[0], 1);
397 /* Skip the size which is encoded int */
398 numa_dist_table += sizeof(__be32);
400 pr_debug("numa_dist_table_len = %d, numa_dist_indexes_len = %d\n",
401 numa_dist_table_length, max_numa_index);
403 for (i = 0; i < max_numa_index; i++)
404 /* +1 skip the max_numa_index in the property */
405 numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
408 if (numa_dist_table_length != max_numa_index * max_numa_index) {
409 WARN(1, "Wrong NUMA distance information\n");
410 /* consider everybody else just remote. */
411 for (i = 0; i < max_numa_index; i++) {
412 for (j = 0; j < max_numa_index; j++) {
413 int nodeA = numa_id_index_table[i];
414 int nodeB = numa_id_index_table[j];
417 numa_distance_table[nodeA][nodeB] = LOCAL_DISTANCE;
419 numa_distance_table[nodeA][nodeB] = REMOTE_DISTANCE;
425 for (i = 0; i < max_numa_index; i++) {
426 for (j = 0; j < max_numa_index; j++) {
427 int nodeA = numa_id_index_table[i];
428 int nodeB = numa_id_index_table[j];
430 numa_distance_table[nodeA][nodeB] = numa_dist_table[distance_index++];
431 pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, numa_distance_table[nodeA][nodeB]);
437 static int __init find_primary_domain_index(void)
440 struct device_node *root;
443 * Check for which form of affinity.
445 if (firmware_has_feature(FW_FEATURE_OPAL)) {
446 affinity_form = FORM1_AFFINITY;
447 } else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
448 pr_debug("Using form 2 affinity\n");
449 affinity_form = FORM2_AFFINITY;
450 } else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
451 pr_debug("Using form 1 affinity\n");
452 affinity_form = FORM1_AFFINITY;
454 affinity_form = FORM0_AFFINITY;
456 if (firmware_has_feature(FW_FEATURE_OPAL))
457 root = of_find_node_by_path("/ibm,opal");
459 root = of_find_node_by_path("/rtas");
461 root = of_find_node_by_path("/");
464 * This property is a set of 32-bit integers, each representing
465 * an index into the ibm,associativity nodes.
467 * With form 0 affinity the first integer is for an SMP configuration
468 * (should be all 0's) and the second is for a normal NUMA
469 * configuration. We have only one level of NUMA.
471 * With form 1 affinity the first integer is the most significant
472 * NUMA boundary and the following are progressively less significant
473 * boundaries. There can be more than one level of NUMA.
475 distance_ref_points = of_get_property(root,
476 "ibm,associativity-reference-points",
477 &distance_ref_points_depth);
479 if (!distance_ref_points) {
480 pr_debug("ibm,associativity-reference-points not found.\n");
484 distance_ref_points_depth /= sizeof(int);
485 if (affinity_form == FORM0_AFFINITY) {
486 if (distance_ref_points_depth < 2) {
487 pr_warn("short ibm,associativity-reference-points\n");
491 index = of_read_number(&distance_ref_points[1], 1);
494 * Both FORM1 and FORM2 affinity find the primary domain details
495 * at the same offset.
497 index = of_read_number(distance_ref_points, 1);
500 * Warn and cap if the hardware supports more than
501 * MAX_DISTANCE_REF_POINTS domains.
503 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
504 pr_warn("distance array capped at %d entries\n",
505 MAX_DISTANCE_REF_POINTS);
506 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
517 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
519 struct device_node *memory = NULL;
521 memory = of_find_node_by_type(memory, "memory");
523 panic("numa.c: No memory nodes found!");
525 *n_addr_cells = of_n_addr_cells(memory);
526 *n_size_cells = of_n_size_cells(memory);
530 static unsigned long read_n_cells(int n, const __be32 **buf)
532 unsigned long result = 0;
535 result = (result << 32) | of_read_number(*buf, 1);
541 struct assoc_arrays {
544 const __be32 *arrays;
548 * Retrieve and validate the list of associativity arrays for drconf
549 * memory from the ibm,associativity-lookup-arrays property of the
552 * The layout of the ibm,associativity-lookup-arrays property is a number N
553 * indicating the number of associativity arrays, followed by a number M
554 * indicating the size of each associativity array, followed by a list
555 * of N associativity arrays.
557 static int of_get_assoc_arrays(struct assoc_arrays *aa)
559 struct device_node *memory;
563 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
567 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
568 if (!prop || len < 2 * sizeof(unsigned int)) {
573 aa->n_arrays = of_read_number(prop++, 1);
574 aa->array_sz = of_read_number(prop++, 1);
578 /* Now that we know the number of arrays and size of each array,
579 * revalidate the size of the property read in.
581 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
588 static int get_nid_and_numa_distance(struct drmem_lmb *lmb)
590 struct assoc_arrays aa = { .arrays = NULL };
591 int default_nid = NUMA_NO_NODE;
592 int nid = default_nid;
595 if ((primary_domain_index < 0) || !numa_enabled)
598 rc = of_get_assoc_arrays(&aa);
602 if (primary_domain_index <= aa.array_sz &&
603 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
604 const __be32 *associativity;
606 index = lmb->aa_index * aa.array_sz;
607 associativity = &aa.arrays[index];
608 nid = __associativity_to_nid(associativity, aa.array_sz);
609 if (nid > 0 && affinity_form == FORM1_AFFINITY) {
611 * lookup array associativity entries have
612 * no length of the array as the first element.
614 __initialize_form1_numa_distance(associativity, aa.array_sz);
621 * This is like of_node_to_nid_single() for memory represented in the
622 * ibm,dynamic-reconfiguration-memory node.
624 int of_drconf_to_nid_single(struct drmem_lmb *lmb)
626 struct assoc_arrays aa = { .arrays = NULL };
627 int default_nid = NUMA_NO_NODE;
628 int nid = default_nid;
631 if ((primary_domain_index < 0) || !numa_enabled)
634 rc = of_get_assoc_arrays(&aa);
638 if (primary_domain_index <= aa.array_sz &&
639 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
640 const __be32 *associativity;
642 index = lmb->aa_index * aa.array_sz;
643 associativity = &aa.arrays[index];
644 nid = __associativity_to_nid(associativity, aa.array_sz);
649 #ifdef CONFIG_PPC_SPLPAR
651 static int __vphn_get_associativity(long lcpu, __be32 *associativity)
656 * On a shared lpar, device tree will not have node associativity.
657 * At this time lppaca, or its __old_status field may not be
658 * updated. Hence kernel cannot detect if its on a shared lpar. So
659 * request an explicit associativity irrespective of whether the
660 * lpar is shared or dedicated. Use the device tree property as a
661 * fallback. cpu_to_phys_id is only valid between
662 * smp_setup_cpu_maps() and smp_setup_pacas().
664 if (firmware_has_feature(FW_FEATURE_VPHN)) {
666 hwid = cpu_to_phys_id[lcpu];
668 hwid = get_hard_smp_processor_id(lcpu);
670 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
678 static int vphn_get_nid(long lcpu)
680 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
683 if (!__vphn_get_associativity(lcpu, associativity))
684 return associativity_to_nid(associativity);
691 static int __vphn_get_associativity(long lcpu, __be32 *associativity)
696 static int vphn_get_nid(long unused)
700 #endif /* CONFIG_PPC_SPLPAR */
703 * Figure out to which domain a cpu belongs and stick it there.
704 * Return the id of the domain used.
706 static int numa_setup_cpu(unsigned long lcpu)
708 struct device_node *cpu;
709 int fcpu = cpu_first_thread_sibling(lcpu);
710 int nid = NUMA_NO_NODE;
712 if (!cpu_present(lcpu)) {
713 set_cpu_numa_node(lcpu, first_online_node);
714 return first_online_node;
718 * If a valid cpu-to-node mapping is already available, use it
719 * directly instead of querying the firmware, since it represents
720 * the most recent mapping notified to us by the platform (eg: VPHN).
721 * Since cpu_to_node binding remains the same for all threads in the
722 * core. If a valid cpu-to-node mapping is already available, for
723 * the first thread in the core, use it.
725 nid = numa_cpu_lookup_table[fcpu];
727 map_cpu_to_node(lcpu, nid);
731 nid = vphn_get_nid(lcpu);
732 if (nid != NUMA_NO_NODE)
735 cpu = of_get_cpu_node(lcpu, NULL);
739 if (cpu_present(lcpu))
745 nid = of_node_to_nid_single(cpu);
749 if (nid < 0 || !node_possible(nid))
750 nid = first_online_node;
753 * Update for the first thread of the core. All threads of a core
754 * have to be part of the same node. This not only avoids querying
755 * for every other thread in the core, but always avoids a case
756 * where virtual node associativity change causes subsequent threads
757 * of a core to be associated with different nid. However if first
758 * thread is already online, expect it to have a valid mapping.
761 WARN_ON(cpu_online(fcpu));
762 map_cpu_to_node(fcpu, nid);
765 map_cpu_to_node(lcpu, nid);
770 static void verify_cpu_node_mapping(int cpu, int node)
772 int base, sibling, i;
774 /* Verify that all the threads in the core belong to the same node */
775 base = cpu_first_thread_sibling(cpu);
777 for (i = 0; i < threads_per_core; i++) {
780 if (sibling == cpu || cpu_is_offline(sibling))
783 if (cpu_to_node(sibling) != node) {
784 WARN(1, "CPU thread siblings %d and %d don't belong"
785 " to the same node!\n", cpu, sibling);
791 /* Must run before sched domains notifier. */
792 static int ppc_numa_cpu_prepare(unsigned int cpu)
796 nid = numa_setup_cpu(cpu);
797 verify_cpu_node_mapping(cpu, nid);
801 static int ppc_numa_cpu_dead(unsigned int cpu)
803 #ifdef CONFIG_HOTPLUG_CPU
804 unmap_cpu_from_node(cpu);
810 * Check and possibly modify a memory region to enforce the memory limit.
812 * Returns the size the region should have to enforce the memory limit.
813 * This will either be the original value of size, a truncated value,
814 * or zero. If the returned value of size is 0 the region should be
815 * discarded as it lies wholly above the memory limit.
817 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
821 * We use memblock_end_of_DRAM() in here instead of memory_limit because
822 * we've already adjusted it for the limit and it takes care of
823 * having memory holes below the limit. Also, in the case of
824 * iommu_is_off, memory_limit is not set but is implicitly enforced.
827 if (start + size <= memblock_end_of_DRAM())
830 if (start >= memblock_end_of_DRAM())
833 return memblock_end_of_DRAM() - start;
837 * Reads the counter for a given entry in
838 * linux,drconf-usable-memory property
840 static inline int __init read_usm_ranges(const __be32 **usm)
843 * For each lmb in ibm,dynamic-memory a corresponding
844 * entry in linux,drconf-usable-memory property contains
845 * a counter followed by that many (base, size) duple.
846 * read the counter from linux,drconf-usable-memory
848 return read_n_cells(n_mem_size_cells, usm);
852 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
853 * node. This assumes n_mem_{addr,size}_cells have been set.
855 static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
859 unsigned int ranges, is_kexec_kdump = 0;
860 unsigned long base, size, sz;
864 * Skip this block if the reserved bit is set in flags (0x80)
865 * or if the block is not assigned to this partition (0x8)
867 if ((lmb->flags & DRCONF_MEM_RESERVED)
868 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
874 base = lmb->base_addr;
875 size = drmem_lmb_size();
878 if (is_kexec_kdump) {
879 ranges = read_usm_ranges(usm);
880 if (!ranges) /* there are no (base, size) duple */
885 if (is_kexec_kdump) {
886 base = read_n_cells(n_mem_addr_cells, usm);
887 size = read_n_cells(n_mem_size_cells, usm);
890 nid = get_nid_and_numa_distance(lmb);
891 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
893 node_set_online(nid);
894 sz = numa_enforce_memory_limit(base, size);
896 memblock_set_node(base, sz, &memblock.memory, nid);
902 static int __init parse_numa_properties(void)
904 struct device_node *memory;
907 const __be32 *associativity;
909 if (numa_enabled == 0) {
910 pr_warn("disabled by user\n");
914 primary_domain_index = find_primary_domain_index();
916 if (primary_domain_index < 0) {
918 * if we fail to parse primary_domain_index from device tree
919 * mark the numa disabled, boot with numa disabled.
921 numa_enabled = false;
922 return primary_domain_index;
925 pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index);
928 * If it is FORM2 initialize the distance table here.
930 if (affinity_form == FORM2_AFFINITY)
931 initialize_form2_numa_distance_lookup_table();
934 * Even though we connect cpus to numa domains later in SMP
935 * init, we need to know the node ids now. This is because
936 * each node to be onlined must have NODE_DATA etc backing it.
938 for_each_present_cpu(i) {
939 __be32 vphn_assoc[VPHN_ASSOC_BUFSIZE];
940 struct device_node *cpu;
941 int nid = NUMA_NO_NODE;
943 memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32));
945 if (__vphn_get_associativity(i, vphn_assoc) == 0) {
946 nid = associativity_to_nid(vphn_assoc);
947 initialize_form1_numa_distance(vphn_assoc);
951 * Don't fall back to default_nid yet -- we will plug
952 * cpus into nodes once the memory scan has discovered
955 cpu = of_get_cpu_node(i, NULL);
958 associativity = of_get_associativity(cpu);
960 nid = associativity_to_nid(associativity);
961 initialize_form1_numa_distance(associativity);
966 node_set_online(nid);
969 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
971 for_each_node_by_type(memory, "memory") {
976 const __be32 *memcell_buf;
979 memcell_buf = of_get_property(memory,
980 "linux,usable-memory", &len);
981 if (!memcell_buf || len <= 0)
982 memcell_buf = of_get_property(memory, "reg", &len);
983 if (!memcell_buf || len <= 0)
987 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
989 /* these are order-sensitive, and modify the buffer pointer */
990 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
991 size = read_n_cells(n_mem_size_cells, &memcell_buf);
994 * Assumption: either all memory nodes or none will
995 * have associativity properties. If none, then
996 * everything goes to default_nid.
998 associativity = of_get_associativity(memory);
1000 nid = associativity_to_nid(associativity);
1001 initialize_form1_numa_distance(associativity);
1005 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
1006 node_set_online(nid);
1008 size = numa_enforce_memory_limit(start, size);
1010 memblock_set_node(start, size, &memblock.memory, nid);
1017 * Now do the same thing for each MEMBLOCK listed in the
1018 * ibm,dynamic-memory property in the
1019 * ibm,dynamic-reconfiguration-memory node.
1021 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1023 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
1024 of_node_put(memory);
1030 static void __init setup_nonnuma(void)
1032 unsigned long top_of_ram = memblock_end_of_DRAM();
1033 unsigned long total_ram = memblock_phys_mem_size();
1034 unsigned long start_pfn, end_pfn;
1035 unsigned int nid = 0;
1038 pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram);
1039 pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20);
1041 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1042 fake_numa_create_new_node(end_pfn, &nid);
1043 memblock_set_node(PFN_PHYS(start_pfn),
1044 PFN_PHYS(end_pfn - start_pfn),
1045 &memblock.memory, nid);
1046 node_set_online(nid);
1050 void __init dump_numa_cpu_topology(void)
1053 unsigned int cpu, count;
1058 for_each_online_node(node) {
1059 pr_info("Node %d CPUs:", node);
1063 * If we used a CPU iterator here we would miss printing
1064 * the holes in the cpumap.
1066 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1067 if (cpumask_test_cpu(cpu,
1068 node_to_cpumask_map[node])) {
1070 pr_cont(" %u", cpu);
1074 pr_cont("-%u", cpu - 1);
1080 pr_cont("-%u", nr_cpu_ids - 1);
1085 /* Initialize NODE_DATA for a node on the local memory */
1086 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
1088 u64 spanned_pages = end_pfn - start_pfn;
1089 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
1094 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
1096 panic("Cannot allocate %zu bytes for node %d data\n",
1101 /* report and initialize */
1102 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
1103 nd_pa, nd_pa + nd_size - 1);
1104 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
1106 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
1108 node_data[nid] = nd;
1109 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
1110 NODE_DATA(nid)->node_id = nid;
1111 NODE_DATA(nid)->node_start_pfn = start_pfn;
1112 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
1115 static void __init find_possible_nodes(void)
1117 struct device_node *rtas;
1118 const __be32 *domains = NULL;
1119 int prop_length, max_nodes;
1125 rtas = of_find_node_by_path("/rtas");
1130 * ibm,current-associativity-domains is a fairly recent property. If
1131 * it doesn't exist, then fallback on ibm,max-associativity-domains.
1132 * Current denotes what the platform can support compared to max
1133 * which denotes what the Hypervisor can support.
1135 * If the LPAR is migratable, new nodes might be activated after a LPM,
1136 * so we should consider the max number in that case.
1138 if (!of_get_property(of_root, "ibm,migratable-partition", NULL))
1139 domains = of_get_property(rtas,
1140 "ibm,current-associativity-domains",
1143 domains = of_get_property(rtas, "ibm,max-associativity-domains",
1149 max_nodes = of_read_number(&domains[primary_domain_index], 1);
1150 pr_info("Partition configured for %d NUMA nodes.\n", max_nodes);
1152 for (i = 0; i < max_nodes; i++) {
1153 if (!node_possible(i))
1154 node_set(i, node_possible_map);
1157 prop_length /= sizeof(int);
1158 if (prop_length > primary_domain_index + 2)
1159 coregroup_enabled = 1;
1165 void __init mem_topology_setup(void)
1170 * Linux/mm assumes node 0 to be online at boot. However this is not
1171 * true on PowerPC, where node 0 is similar to any other node, it
1172 * could be cpuless, memoryless node. So force node 0 to be offline
1173 * for now. This will prevent cpuless, memoryless node 0 showing up
1174 * unnecessarily as online. If a node has cpus or memory that need
1175 * to be online, then node will anyway be marked online.
1177 node_set_offline(0);
1179 if (parse_numa_properties())
1183 * Modify the set of possible NUMA nodes to reflect information
1184 * available about the set of online nodes, and the set of nodes
1185 * that we expect to make use of for this platform's affinity
1188 nodes_and(node_possible_map, node_possible_map, node_online_map);
1190 find_possible_nodes();
1192 setup_node_to_cpumask_map();
1194 reset_numa_cpu_lookup_table();
1196 for_each_possible_cpu(cpu) {
1198 * Powerpc with CONFIG_NUMA always used to have a node 0,
1199 * even if it was memoryless or cpuless. For all cpus that
1200 * are possible but not present, cpu_to_node() would point
1201 * to node 0. To remove a cpuless, memoryless dummy node,
1202 * powerpc need to make sure all possible but not present
1203 * cpu_to_node are set to a proper node.
1205 numa_setup_cpu(cpu);
1209 void __init initmem_init(void)
1213 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1214 max_pfn = max_low_pfn;
1216 memblock_dump_all();
1218 for_each_online_node(nid) {
1219 unsigned long start_pfn, end_pfn;
1221 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1222 setup_node_data(nid, start_pfn, end_pfn);
1228 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1229 * even before we online them, so that we can use cpu_to_{node,mem}
1230 * early in boot, cf. smp_prepare_cpus().
1231 * _nocalls() + manual invocation is used because cpuhp is not yet
1232 * initialized for the boot CPU.
1234 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1235 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
1238 static int __init early_numa(char *p)
1243 if (strstr(p, "off"))
1246 p = strstr(p, "fake=");
1248 cmdline = p + strlen("fake=");
1252 early_param("numa", early_numa);
1254 #ifdef CONFIG_MEMORY_HOTPLUG
1256 * Find the node associated with a hot added memory section for
1257 * memory represented in the device tree by the property
1258 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1260 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
1262 struct drmem_lmb *lmb;
1263 unsigned long lmb_size;
1264 int nid = NUMA_NO_NODE;
1266 lmb_size = drmem_lmb_size();
1268 for_each_drmem_lmb(lmb) {
1269 /* skip this block if it is reserved or not assigned to
1271 if ((lmb->flags & DRCONF_MEM_RESERVED)
1272 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1275 if ((scn_addr < lmb->base_addr)
1276 || (scn_addr >= (lmb->base_addr + lmb_size)))
1279 nid = of_drconf_to_nid_single(lmb);
1287 * Find the node associated with a hot added memory section for memory
1288 * represented in the device tree as a node (i.e. memory@XXXX) for
1291 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1293 struct device_node *memory;
1294 int nid = NUMA_NO_NODE;
1296 for_each_node_by_type(memory, "memory") {
1297 unsigned long start, size;
1299 const __be32 *memcell_buf;
1302 memcell_buf = of_get_property(memory, "reg", &len);
1303 if (!memcell_buf || len <= 0)
1306 /* ranges in cell */
1307 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1310 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1311 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1313 if ((scn_addr < start) || (scn_addr >= (start + size)))
1316 nid = of_node_to_nid_single(memory);
1324 of_node_put(memory);
1330 * Find the node associated with a hot added memory section. Section
1331 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1332 * sections are fully contained within a single MEMBLOCK.
1334 int hot_add_scn_to_nid(unsigned long scn_addr)
1336 struct device_node *memory = NULL;
1340 return first_online_node;
1342 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1344 nid = hot_add_drconf_scn_to_nid(scn_addr);
1345 of_node_put(memory);
1347 nid = hot_add_node_scn_to_nid(scn_addr);
1350 if (nid < 0 || !node_possible(nid))
1351 nid = first_online_node;
1356 static u64 hot_add_drconf_memory_max(void)
1358 struct device_node *memory = NULL;
1359 struct device_node *dn = NULL;
1360 const __be64 *lrdr = NULL;
1362 dn = of_find_node_by_path("/rtas");
1364 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1367 return be64_to_cpup(lrdr);
1370 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1372 of_node_put(memory);
1373 return drmem_lmb_memory_max();
1379 * memory_hotplug_max - return max address of memory that may be added
1381 * This is currently only used on systems that support drconfig memory
1384 u64 memory_hotplug_max(void)
1386 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1388 #endif /* CONFIG_MEMORY_HOTPLUG */
1390 /* Virtual Processor Home Node (VPHN) support */
1391 #ifdef CONFIG_PPC_SPLPAR
1392 static int topology_inited;
1395 * Retrieve the new associativity information for a virtual processor's
1398 static long vphn_get_associativity(unsigned long cpu,
1399 __be32 *associativity)
1403 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1404 VPHN_FLAG_VCPU, associativity);
1408 pr_debug("VPHN hcall succeeded. Reset polling...\n");
1412 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
1415 pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
1416 "preventing VPHN. Disabling polling...\n");
1419 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1420 "Disabling polling...\n");
1423 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1431 int find_and_online_cpu_nid(int cpu)
1433 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1436 /* Use associativity from first thread for all siblings */
1437 if (vphn_get_associativity(cpu, associativity))
1438 return cpu_to_node(cpu);
1440 new_nid = associativity_to_nid(associativity);
1441 if (new_nid < 0 || !node_possible(new_nid))
1442 new_nid = first_online_node;
1444 if (NODE_DATA(new_nid) == NULL) {
1445 #ifdef CONFIG_MEMORY_HOTPLUG
1447 * Need to ensure that NODE_DATA is initialized for a node from
1448 * available memory (see memblock_alloc_try_nid). If unable to
1449 * init the node, then default to nearest node that has memory
1450 * installed. Skip onlining a node if the subsystems are not
1453 if (!topology_inited || try_online_node(new_nid))
1454 new_nid = first_online_node;
1457 * Default to using the nearest node that has memory installed.
1458 * Otherwise, it would be necessary to patch the kernel MM code
1459 * to deal with more memoryless-node error conditions.
1461 new_nid = first_online_node;
1465 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1470 int cpu_to_coregroup_id(int cpu)
1472 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1475 if (cpu < 0 || cpu > nr_cpu_ids)
1478 if (!coregroup_enabled)
1481 if (!firmware_has_feature(FW_FEATURE_VPHN))
1484 if (vphn_get_associativity(cpu, associativity))
1487 index = of_read_number(associativity, 1);
1488 if (index > primary_domain_index + 1)
1489 return of_read_number(&associativity[index - 1], 1);
1492 return cpu_to_core_id(cpu);
1495 static int topology_update_init(void)
1497 topology_inited = 1;
1500 device_initcall(topology_update_init);
1501 #endif /* CONFIG_PPC_SPLPAR */