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>
30 #include <asm/topology.h>
31 #include <asm/firmware.h>
33 #include <asm/hvcall.h>
34 #include <asm/setup.h>
36 #include <asm/drmem.h>
38 static int numa_enabled = 1;
40 static char *cmdline __initdata;
42 int numa_cpu_lookup_table[NR_CPUS];
43 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
44 struct pglist_data *node_data[MAX_NUMNODES];
46 EXPORT_SYMBOL(numa_cpu_lookup_table);
47 EXPORT_SYMBOL(node_to_cpumask_map);
48 EXPORT_SYMBOL(node_data);
50 static int primary_domain_index;
51 static int n_mem_addr_cells, n_mem_size_cells;
53 #define FORM0_AFFINITY 0
54 #define FORM1_AFFINITY 1
55 #define FORM2_AFFINITY 2
56 static int affinity_form;
58 #define MAX_DISTANCE_REF_POINTS 4
59 static int distance_ref_points_depth;
60 static const __be32 *distance_ref_points;
61 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
62 static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
63 [0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
65 static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
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 pr_debug("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 pr_debug("created new fake_node with id %d\n", fake_nid);
136 static void __init reset_numa_cpu_lookup_table(void)
140 for_each_possible_cpu(cpu)
141 numa_cpu_lookup_table[cpu] = -1;
144 void map_cpu_to_node(int cpu, int node)
146 update_numa_cpu_lookup_table(cpu, node);
148 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) {
149 pr_debug("adding cpu %d to node %d\n", cpu, node);
150 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
154 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
155 void unmap_cpu_from_node(unsigned long cpu)
157 int node = numa_cpu_lookup_table[cpu];
159 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
160 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
161 pr_debug("removing cpu %lu from node %d\n", cpu, node);
163 pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node);
166 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
168 static int __associativity_to_nid(const __be32 *associativity,
173 * primary_domain_index is 1 based array index.
175 int index = primary_domain_index - 1;
177 if (!numa_enabled || index >= max_array_sz)
180 nid = of_read_number(&associativity[index], 1);
182 /* POWER4 LPAR uses 0xffff as invalid node */
183 if (nid == 0xffff || nid >= nr_node_ids)
188 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
191 static int associativity_to_nid(const __be32 *associativity)
193 int array_sz = of_read_number(associativity, 1);
195 /* Skip the first element in the associativity array */
196 return __associativity_to_nid((associativity + 1), array_sz);
199 static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
204 node1 = associativity_to_nid(cpu1_assoc);
205 node2 = associativity_to_nid(cpu2_assoc);
207 dist = numa_distance_table[node1][node2];
208 if (dist <= LOCAL_DISTANCE)
210 else if (dist <= REMOTE_DISTANCE)
216 static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
222 for (i = 0; i < distance_ref_points_depth; i++) {
223 index = be32_to_cpu(distance_ref_points[i]);
224 if (cpu1_assoc[index] == cpu2_assoc[index])
232 int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
234 /* We should not get called with FORM0 */
235 VM_WARN_ON(affinity_form == FORM0_AFFINITY);
236 if (affinity_form == FORM1_AFFINITY)
237 return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
238 return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
241 /* must hold reference to node during call */
242 static const __be32 *of_get_associativity(struct device_node *dev)
244 return of_get_property(dev, "ibm,associativity", NULL);
247 int __node_distance(int a, int b)
250 int distance = LOCAL_DISTANCE;
252 if (affinity_form == FORM2_AFFINITY)
253 return numa_distance_table[a][b];
254 else if (affinity_form == FORM0_AFFINITY)
255 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
257 for (i = 0; i < distance_ref_points_depth; i++) {
258 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
261 /* Double the distance for each NUMA level */
267 EXPORT_SYMBOL(__node_distance);
269 /* Returns the nid associated with the given device tree node,
270 * or -1 if not found.
272 static int of_node_to_nid_single(struct device_node *device)
274 int nid = NUMA_NO_NODE;
277 tmp = of_get_associativity(device);
279 nid = associativity_to_nid(tmp);
283 /* Walk the device tree upwards, looking for an associativity id */
284 int of_node_to_nid(struct device_node *device)
286 int nid = NUMA_NO_NODE;
290 nid = of_node_to_nid_single(device);
294 device = of_get_next_parent(device);
300 EXPORT_SYMBOL(of_node_to_nid);
302 static void __initialize_form1_numa_distance(const __be32 *associativity,
307 if (affinity_form != FORM1_AFFINITY)
310 nid = __associativity_to_nid(associativity, max_array_sz);
311 if (nid != NUMA_NO_NODE) {
312 for (i = 0; i < distance_ref_points_depth; i++) {
314 int index = be32_to_cpu(distance_ref_points[i]) - 1;
317 * broken hierarchy, return with broken distance table
319 if (WARN(index >= max_array_sz, "Broken ibm,associativity property"))
322 entry = &associativity[index];
323 distance_lookup_table[nid][i] = of_read_number(entry, 1);
328 static void initialize_form1_numa_distance(const __be32 *associativity)
332 array_sz = of_read_number(associativity, 1);
333 /* Skip the first element in the associativity array */
334 __initialize_form1_numa_distance(associativity + 1, array_sz);
338 * Used to update distance information w.r.t newly added node.
340 void update_numa_distance(struct device_node *node)
344 if (affinity_form == FORM0_AFFINITY)
346 else if (affinity_form == FORM1_AFFINITY) {
347 const __be32 *associativity;
349 associativity = of_get_associativity(node);
353 initialize_form1_numa_distance(associativity);
358 nid = of_node_to_nid_single(node);
359 if (nid == NUMA_NO_NODE)
363 * With FORM2 we expect NUMA distance of all possible NUMA
364 * nodes to be provided during boot.
366 WARN(numa_distance_table[nid][nid] == -1,
367 "NUMA distance details for node %d not provided\n", nid);
371 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
372 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
374 static void __init initialize_form2_numa_distance_lookup_table(void)
377 struct device_node *root;
378 const __u8 *form2_distances;
379 const __be32 *numa_lookup_index;
380 int form2_distances_length;
381 int max_numa_index, distance_index;
383 if (firmware_has_feature(FW_FEATURE_OPAL))
384 root = of_find_node_by_path("/ibm,opal");
386 root = of_find_node_by_path("/rtas");
388 root = of_find_node_by_path("/");
390 numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
391 max_numa_index = of_read_number(&numa_lookup_index[0], 1);
393 /* first element of the array is the size and is encode-int */
394 form2_distances = of_get_property(root, "ibm,numa-distance-table", NULL);
395 form2_distances_length = of_read_number((const __be32 *)&form2_distances[0], 1);
396 /* Skip the size which is encoded int */
397 form2_distances += sizeof(__be32);
399 pr_debug("form2_distances_len = %d, numa_dist_indexes_len = %d\n",
400 form2_distances_length, max_numa_index);
402 for (i = 0; i < max_numa_index; i++)
403 /* +1 skip the max_numa_index in the property */
404 numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
407 if (form2_distances_length != max_numa_index * max_numa_index) {
408 WARN(1, "Wrong NUMA distance information\n");
409 form2_distances = NULL; // don't use it
412 for (i = 0; i < max_numa_index; i++) {
413 for (j = 0; j < max_numa_index; j++) {
414 int nodeA = numa_id_index_table[i];
415 int nodeB = numa_id_index_table[j];
419 dist = form2_distances[distance_index++];
420 else if (nodeA == nodeB)
421 dist = LOCAL_DISTANCE;
423 dist = REMOTE_DISTANCE;
424 numa_distance_table[nodeA][nodeB] = dist;
425 pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, dist);
432 static int __init find_primary_domain_index(void)
435 struct device_node *root;
438 * Check for which form of affinity.
440 if (firmware_has_feature(FW_FEATURE_OPAL)) {
441 affinity_form = FORM1_AFFINITY;
442 } else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
443 pr_debug("Using form 2 affinity\n");
444 affinity_form = FORM2_AFFINITY;
445 } else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
446 pr_debug("Using form 1 affinity\n");
447 affinity_form = FORM1_AFFINITY;
449 affinity_form = FORM0_AFFINITY;
451 if (firmware_has_feature(FW_FEATURE_OPAL))
452 root = of_find_node_by_path("/ibm,opal");
454 root = of_find_node_by_path("/rtas");
456 root = of_find_node_by_path("/");
459 * This property is a set of 32-bit integers, each representing
460 * an index into the ibm,associativity nodes.
462 * With form 0 affinity the first integer is for an SMP configuration
463 * (should be all 0's) and the second is for a normal NUMA
464 * configuration. We have only one level of NUMA.
466 * With form 1 affinity the first integer is the most significant
467 * NUMA boundary and the following are progressively less significant
468 * boundaries. There can be more than one level of NUMA.
470 distance_ref_points = of_get_property(root,
471 "ibm,associativity-reference-points",
472 &distance_ref_points_depth);
474 if (!distance_ref_points) {
475 pr_debug("ibm,associativity-reference-points not found.\n");
479 distance_ref_points_depth /= sizeof(int);
480 if (affinity_form == FORM0_AFFINITY) {
481 if (distance_ref_points_depth < 2) {
482 pr_warn("short ibm,associativity-reference-points\n");
486 index = of_read_number(&distance_ref_points[1], 1);
489 * Both FORM1 and FORM2 affinity find the primary domain details
490 * at the same offset.
492 index = of_read_number(distance_ref_points, 1);
495 * Warn and cap if the hardware supports more than
496 * MAX_DISTANCE_REF_POINTS domains.
498 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
499 pr_warn("distance array capped at %d entries\n",
500 MAX_DISTANCE_REF_POINTS);
501 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
512 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
514 struct device_node *memory = NULL;
516 memory = of_find_node_by_type(memory, "memory");
518 panic("numa.c: No memory nodes found!");
520 *n_addr_cells = of_n_addr_cells(memory);
521 *n_size_cells = of_n_size_cells(memory);
525 static unsigned long read_n_cells(int n, const __be32 **buf)
527 unsigned long result = 0;
530 result = (result << 32) | of_read_number(*buf, 1);
536 struct assoc_arrays {
539 const __be32 *arrays;
543 * Retrieve and validate the list of associativity arrays for drconf
544 * memory from the ibm,associativity-lookup-arrays property of the
547 * The layout of the ibm,associativity-lookup-arrays property is a number N
548 * indicating the number of associativity arrays, followed by a number M
549 * indicating the size of each associativity array, followed by a list
550 * of N associativity arrays.
552 static int of_get_assoc_arrays(struct assoc_arrays *aa)
554 struct device_node *memory;
558 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
562 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
563 if (!prop || len < 2 * sizeof(unsigned int)) {
568 aa->n_arrays = of_read_number(prop++, 1);
569 aa->array_sz = of_read_number(prop++, 1);
573 /* Now that we know the number of arrays and size of each array,
574 * revalidate the size of the property read in.
576 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
583 static int __init get_nid_and_numa_distance(struct drmem_lmb *lmb)
585 struct assoc_arrays aa = { .arrays = NULL };
586 int default_nid = NUMA_NO_NODE;
587 int nid = default_nid;
590 if ((primary_domain_index < 0) || !numa_enabled)
593 rc = of_get_assoc_arrays(&aa);
597 if (primary_domain_index <= aa.array_sz &&
598 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
599 const __be32 *associativity;
601 index = lmb->aa_index * aa.array_sz;
602 associativity = &aa.arrays[index];
603 nid = __associativity_to_nid(associativity, aa.array_sz);
604 if (nid > 0 && affinity_form == FORM1_AFFINITY) {
606 * lookup array associativity entries have
607 * no length of the array as the first element.
609 __initialize_form1_numa_distance(associativity, aa.array_sz);
616 * This is like of_node_to_nid_single() for memory represented in the
617 * ibm,dynamic-reconfiguration-memory node.
619 int of_drconf_to_nid_single(struct drmem_lmb *lmb)
621 struct assoc_arrays aa = { .arrays = NULL };
622 int default_nid = NUMA_NO_NODE;
623 int nid = default_nid;
626 if ((primary_domain_index < 0) || !numa_enabled)
629 rc = of_get_assoc_arrays(&aa);
633 if (primary_domain_index <= aa.array_sz &&
634 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
635 const __be32 *associativity;
637 index = lmb->aa_index * aa.array_sz;
638 associativity = &aa.arrays[index];
639 nid = __associativity_to_nid(associativity, aa.array_sz);
644 #ifdef CONFIG_PPC_SPLPAR
646 static int __vphn_get_associativity(long lcpu, __be32 *associativity)
651 * On a shared lpar, device tree will not have node associativity.
652 * At this time lppaca, or its __old_status field may not be
653 * updated. Hence kernel cannot detect if its on a shared lpar. So
654 * request an explicit associativity irrespective of whether the
655 * lpar is shared or dedicated. Use the device tree property as a
656 * fallback. cpu_to_phys_id is only valid between
657 * smp_setup_cpu_maps() and smp_setup_pacas().
659 if (firmware_has_feature(FW_FEATURE_VPHN)) {
661 hwid = cpu_to_phys_id[lcpu];
663 hwid = get_hard_smp_processor_id(lcpu);
665 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
673 static int vphn_get_nid(long lcpu)
675 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
678 if (!__vphn_get_associativity(lcpu, associativity))
679 return associativity_to_nid(associativity);
686 static int __vphn_get_associativity(long lcpu, __be32 *associativity)
691 static int vphn_get_nid(long unused)
695 #endif /* CONFIG_PPC_SPLPAR */
698 * Figure out to which domain a cpu belongs and stick it there.
699 * Return the id of the domain used.
701 static int numa_setup_cpu(unsigned long lcpu)
703 struct device_node *cpu;
704 int fcpu = cpu_first_thread_sibling(lcpu);
705 int nid = NUMA_NO_NODE;
707 if (!cpu_present(lcpu)) {
708 set_cpu_numa_node(lcpu, first_online_node);
709 return first_online_node;
713 * If a valid cpu-to-node mapping is already available, use it
714 * directly instead of querying the firmware, since it represents
715 * the most recent mapping notified to us by the platform (eg: VPHN).
716 * Since cpu_to_node binding remains the same for all threads in the
717 * core. If a valid cpu-to-node mapping is already available, for
718 * the first thread in the core, use it.
720 nid = numa_cpu_lookup_table[fcpu];
722 map_cpu_to_node(lcpu, nid);
726 nid = vphn_get_nid(lcpu);
727 if (nid != NUMA_NO_NODE)
730 cpu = of_get_cpu_node(lcpu, NULL);
734 if (cpu_present(lcpu))
740 nid = of_node_to_nid_single(cpu);
744 if (nid < 0 || !node_possible(nid))
745 nid = first_online_node;
748 * Update for the first thread of the core. All threads of a core
749 * have to be part of the same node. This not only avoids querying
750 * for every other thread in the core, but always avoids a case
751 * where virtual node associativity change causes subsequent threads
752 * of a core to be associated with different nid. However if first
753 * thread is already online, expect it to have a valid mapping.
756 WARN_ON(cpu_online(fcpu));
757 map_cpu_to_node(fcpu, nid);
760 map_cpu_to_node(lcpu, nid);
765 static void verify_cpu_node_mapping(int cpu, int node)
767 int base, sibling, i;
769 /* Verify that all the threads in the core belong to the same node */
770 base = cpu_first_thread_sibling(cpu);
772 for (i = 0; i < threads_per_core; i++) {
775 if (sibling == cpu || cpu_is_offline(sibling))
778 if (cpu_to_node(sibling) != node) {
779 WARN(1, "CPU thread siblings %d and %d don't belong"
780 " to the same node!\n", cpu, sibling);
786 /* Must run before sched domains notifier. */
787 static int ppc_numa_cpu_prepare(unsigned int cpu)
791 nid = numa_setup_cpu(cpu);
792 verify_cpu_node_mapping(cpu, nid);
796 static int ppc_numa_cpu_dead(unsigned int cpu)
802 * Check and possibly modify a memory region to enforce the memory limit.
804 * Returns the size the region should have to enforce the memory limit.
805 * This will either be the original value of size, a truncated value,
806 * or zero. If the returned value of size is 0 the region should be
807 * discarded as it lies wholly above the memory limit.
809 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
813 * We use memblock_end_of_DRAM() in here instead of memory_limit because
814 * we've already adjusted it for the limit and it takes care of
815 * having memory holes below the limit. Also, in the case of
816 * iommu_is_off, memory_limit is not set but is implicitly enforced.
819 if (start + size <= memblock_end_of_DRAM())
822 if (start >= memblock_end_of_DRAM())
825 return memblock_end_of_DRAM() - start;
829 * Reads the counter for a given entry in
830 * linux,drconf-usable-memory property
832 static inline int __init read_usm_ranges(const __be32 **usm)
835 * For each lmb in ibm,dynamic-memory a corresponding
836 * entry in linux,drconf-usable-memory property contains
837 * a counter followed by that many (base, size) duple.
838 * read the counter from linux,drconf-usable-memory
840 return read_n_cells(n_mem_size_cells, usm);
844 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
845 * node. This assumes n_mem_{addr,size}_cells have been set.
847 static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
851 unsigned int ranges, is_kexec_kdump = 0;
852 unsigned long base, size, sz;
856 * Skip this block if the reserved bit is set in flags (0x80)
857 * or if the block is not assigned to this partition (0x8)
859 if ((lmb->flags & DRCONF_MEM_RESERVED)
860 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
866 base = lmb->base_addr;
867 size = drmem_lmb_size();
870 if (is_kexec_kdump) {
871 ranges = read_usm_ranges(usm);
872 if (!ranges) /* there are no (base, size) duple */
877 if (is_kexec_kdump) {
878 base = read_n_cells(n_mem_addr_cells, usm);
879 size = read_n_cells(n_mem_size_cells, usm);
882 nid = get_nid_and_numa_distance(lmb);
883 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
885 node_set_online(nid);
886 sz = numa_enforce_memory_limit(base, size);
888 memblock_set_node(base, sz, &memblock.memory, nid);
894 static int __init parse_numa_properties(void)
896 struct device_node *memory;
899 const __be32 *associativity;
901 if (numa_enabled == 0) {
902 pr_warn("disabled by user\n");
906 primary_domain_index = find_primary_domain_index();
908 if (primary_domain_index < 0) {
910 * if we fail to parse primary_domain_index from device tree
911 * mark the numa disabled, boot with numa disabled.
913 numa_enabled = false;
914 return primary_domain_index;
917 pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index);
920 * If it is FORM2 initialize the distance table here.
922 if (affinity_form == FORM2_AFFINITY)
923 initialize_form2_numa_distance_lookup_table();
926 * Even though we connect cpus to numa domains later in SMP
927 * init, we need to know the node ids now. This is because
928 * each node to be onlined must have NODE_DATA etc backing it.
930 for_each_present_cpu(i) {
931 __be32 vphn_assoc[VPHN_ASSOC_BUFSIZE];
932 struct device_node *cpu;
933 int nid = NUMA_NO_NODE;
935 memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32));
937 if (__vphn_get_associativity(i, vphn_assoc) == 0) {
938 nid = associativity_to_nid(vphn_assoc);
939 initialize_form1_numa_distance(vphn_assoc);
943 * Don't fall back to default_nid yet -- we will plug
944 * cpus into nodes once the memory scan has discovered
947 cpu = of_get_cpu_node(i, NULL);
950 associativity = of_get_associativity(cpu);
952 nid = associativity_to_nid(associativity);
953 initialize_form1_numa_distance(associativity);
958 /* node_set_online() is an UB if 'nid' is negative */
959 if (likely(nid >= 0))
960 node_set_online(nid);
963 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
965 for_each_node_by_type(memory, "memory") {
970 const __be32 *memcell_buf;
973 memcell_buf = of_get_property(memory,
974 "linux,usable-memory", &len);
975 if (!memcell_buf || len <= 0)
976 memcell_buf = of_get_property(memory, "reg", &len);
977 if (!memcell_buf || len <= 0)
981 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
983 /* these are order-sensitive, and modify the buffer pointer */
984 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
985 size = read_n_cells(n_mem_size_cells, &memcell_buf);
988 * Assumption: either all memory nodes or none will
989 * have associativity properties. If none, then
990 * everything goes to default_nid.
992 associativity = of_get_associativity(memory);
994 nid = associativity_to_nid(associativity);
995 initialize_form1_numa_distance(associativity);
999 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
1000 node_set_online(nid);
1002 size = numa_enforce_memory_limit(start, size);
1004 memblock_set_node(start, size, &memblock.memory, nid);
1011 * Now do the same thing for each MEMBLOCK listed in the
1012 * ibm,dynamic-memory property in the
1013 * ibm,dynamic-reconfiguration-memory node.
1015 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1017 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
1018 of_node_put(memory);
1024 static void __init setup_nonnuma(void)
1026 unsigned long top_of_ram = memblock_end_of_DRAM();
1027 unsigned long total_ram = memblock_phys_mem_size();
1028 unsigned long start_pfn, end_pfn;
1029 unsigned int nid = 0;
1032 pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram);
1033 pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20);
1035 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1036 fake_numa_create_new_node(end_pfn, &nid);
1037 memblock_set_node(PFN_PHYS(start_pfn),
1038 PFN_PHYS(end_pfn - start_pfn),
1039 &memblock.memory, nid);
1040 node_set_online(nid);
1044 void __init dump_numa_cpu_topology(void)
1047 unsigned int cpu, count;
1052 for_each_online_node(node) {
1053 pr_info("Node %d CPUs:", node);
1057 * If we used a CPU iterator here we would miss printing
1058 * the holes in the cpumap.
1060 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1061 if (cpumask_test_cpu(cpu,
1062 node_to_cpumask_map[node])) {
1064 pr_cont(" %u", cpu);
1068 pr_cont("-%u", cpu - 1);
1074 pr_cont("-%u", nr_cpu_ids - 1);
1079 /* Initialize NODE_DATA for a node on the local memory */
1080 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
1082 u64 spanned_pages = end_pfn - start_pfn;
1083 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
1088 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
1090 panic("Cannot allocate %zu bytes for node %d data\n",
1095 /* report and initialize */
1096 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n",
1097 nd_pa, nd_pa + nd_size - 1);
1098 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
1100 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid);
1102 node_data[nid] = nd;
1103 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
1104 NODE_DATA(nid)->node_id = nid;
1105 NODE_DATA(nid)->node_start_pfn = start_pfn;
1106 NODE_DATA(nid)->node_spanned_pages = spanned_pages;
1109 static void __init find_possible_nodes(void)
1111 struct device_node *rtas;
1112 const __be32 *domains = NULL;
1113 int prop_length, max_nodes;
1119 rtas = of_find_node_by_path("/rtas");
1124 * ibm,current-associativity-domains is a fairly recent property. If
1125 * it doesn't exist, then fallback on ibm,max-associativity-domains.
1126 * Current denotes what the platform can support compared to max
1127 * which denotes what the Hypervisor can support.
1129 * If the LPAR is migratable, new nodes might be activated after a LPM,
1130 * so we should consider the max number in that case.
1132 if (!of_get_property(of_root, "ibm,migratable-partition", NULL))
1133 domains = of_get_property(rtas,
1134 "ibm,current-associativity-domains",
1137 domains = of_get_property(rtas, "ibm,max-associativity-domains",
1143 max_nodes = of_read_number(&domains[primary_domain_index], 1);
1144 pr_info("Partition configured for %d NUMA nodes.\n", max_nodes);
1146 for (i = 0; i < max_nodes; i++) {
1147 if (!node_possible(i))
1148 node_set(i, node_possible_map);
1151 prop_length /= sizeof(int);
1152 if (prop_length > primary_domain_index + 2)
1153 coregroup_enabled = 1;
1159 void __init mem_topology_setup(void)
1164 * Linux/mm assumes node 0 to be online at boot. However this is not
1165 * true on PowerPC, where node 0 is similar to any other node, it
1166 * could be cpuless, memoryless node. So force node 0 to be offline
1167 * for now. This will prevent cpuless, memoryless node 0 showing up
1168 * unnecessarily as online. If a node has cpus or memory that need
1169 * to be online, then node will anyway be marked online.
1171 node_set_offline(0);
1173 if (parse_numa_properties())
1177 * Modify the set of possible NUMA nodes to reflect information
1178 * available about the set of online nodes, and the set of nodes
1179 * that we expect to make use of for this platform's affinity
1182 nodes_and(node_possible_map, node_possible_map, node_online_map);
1184 find_possible_nodes();
1186 setup_node_to_cpumask_map();
1188 reset_numa_cpu_lookup_table();
1190 for_each_possible_cpu(cpu) {
1192 * Powerpc with CONFIG_NUMA always used to have a node 0,
1193 * even if it was memoryless or cpuless. For all cpus that
1194 * are possible but not present, cpu_to_node() would point
1195 * to node 0. To remove a cpuless, memoryless dummy node,
1196 * powerpc need to make sure all possible but not present
1197 * cpu_to_node are set to a proper node.
1199 numa_setup_cpu(cpu);
1203 void __init initmem_init(void)
1207 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1208 max_pfn = max_low_pfn;
1210 memblock_dump_all();
1212 for_each_online_node(nid) {
1213 unsigned long start_pfn, end_pfn;
1215 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1216 setup_node_data(nid, start_pfn, end_pfn);
1222 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1223 * even before we online them, so that we can use cpu_to_{node,mem}
1224 * early in boot, cf. smp_prepare_cpus().
1225 * _nocalls() + manual invocation is used because cpuhp is not yet
1226 * initialized for the boot CPU.
1228 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1229 ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
1232 static int __init early_numa(char *p)
1237 if (strstr(p, "off"))
1240 p = strstr(p, "fake=");
1242 cmdline = p + strlen("fake=");
1246 early_param("numa", early_numa);
1248 #ifdef CONFIG_MEMORY_HOTPLUG
1250 * Find the node associated with a hot added memory section for
1251 * memory represented in the device tree by the property
1252 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1254 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
1256 struct drmem_lmb *lmb;
1257 unsigned long lmb_size;
1258 int nid = NUMA_NO_NODE;
1260 lmb_size = drmem_lmb_size();
1262 for_each_drmem_lmb(lmb) {
1263 /* skip this block if it is reserved or not assigned to
1265 if ((lmb->flags & DRCONF_MEM_RESERVED)
1266 || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1269 if ((scn_addr < lmb->base_addr)
1270 || (scn_addr >= (lmb->base_addr + lmb_size)))
1273 nid = of_drconf_to_nid_single(lmb);
1281 * Find the node associated with a hot added memory section for memory
1282 * represented in the device tree as a node (i.e. memory@XXXX) for
1285 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1287 struct device_node *memory;
1288 int nid = NUMA_NO_NODE;
1290 for_each_node_by_type(memory, "memory") {
1291 unsigned long start, size;
1293 const __be32 *memcell_buf;
1296 memcell_buf = of_get_property(memory, "reg", &len);
1297 if (!memcell_buf || len <= 0)
1300 /* ranges in cell */
1301 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1304 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1305 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1307 if ((scn_addr < start) || (scn_addr >= (start + size)))
1310 nid = of_node_to_nid_single(memory);
1318 of_node_put(memory);
1324 * Find the node associated with a hot added memory section. Section
1325 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1326 * sections are fully contained within a single MEMBLOCK.
1328 int hot_add_scn_to_nid(unsigned long scn_addr)
1330 struct device_node *memory = NULL;
1334 return first_online_node;
1336 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1338 nid = hot_add_drconf_scn_to_nid(scn_addr);
1339 of_node_put(memory);
1341 nid = hot_add_node_scn_to_nid(scn_addr);
1344 if (nid < 0 || !node_possible(nid))
1345 nid = first_online_node;
1350 static u64 hot_add_drconf_memory_max(void)
1352 struct device_node *memory = NULL;
1353 struct device_node *dn = NULL;
1354 const __be64 *lrdr = NULL;
1356 dn = of_find_node_by_path("/rtas");
1358 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1361 return be64_to_cpup(lrdr);
1364 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1366 of_node_put(memory);
1367 return drmem_lmb_memory_max();
1373 * memory_hotplug_max - return max address of memory that may be added
1375 * This is currently only used on systems that support drconfig memory
1378 u64 memory_hotplug_max(void)
1380 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1382 #endif /* CONFIG_MEMORY_HOTPLUG */
1384 /* Virtual Processor Home Node (VPHN) support */
1385 #ifdef CONFIG_PPC_SPLPAR
1386 static int topology_inited;
1389 * Retrieve the new associativity information for a virtual processor's
1392 static long vphn_get_associativity(unsigned long cpu,
1393 __be32 *associativity)
1397 rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1398 VPHN_FLAG_VCPU, associativity);
1402 pr_debug("VPHN hcall succeeded. Reset polling...\n");
1406 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
1409 pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
1410 "preventing VPHN. Disabling polling...\n");
1413 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1414 "Disabling polling...\n");
1417 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1425 void find_and_update_cpu_nid(int cpu)
1427 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1430 /* Use associativity from first thread for all siblings */
1431 if (vphn_get_associativity(cpu, associativity))
1434 /* Do not have previous associativity, so find it now. */
1435 new_nid = associativity_to_nid(associativity);
1437 if (new_nid < 0 || !node_possible(new_nid))
1438 new_nid = first_online_node;
1440 // Associate node <-> cpu, so cpu_up() calls
1441 // try_online_node() on the right node.
1442 set_cpu_numa_node(cpu, new_nid);
1444 pr_debug("%s:%d cpu %d nid %d\n", __func__, __LINE__, cpu, new_nid);
1447 int cpu_to_coregroup_id(int cpu)
1449 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1452 if (cpu < 0 || cpu > nr_cpu_ids)
1455 if (!coregroup_enabled)
1458 if (!firmware_has_feature(FW_FEATURE_VPHN))
1461 if (vphn_get_associativity(cpu, associativity))
1464 index = of_read_number(associativity, 1);
1465 if (index > primary_domain_index + 1)
1466 return of_read_number(&associativity[index - 1], 1);
1469 return cpu_to_core_id(cpu);
1472 static int topology_update_init(void)
1474 topology_inited = 1;
1477 device_initcall(topology_update_init);
1478 #endif /* CONFIG_PPC_SPLPAR */