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 #include <linux/threads.h>
12 #include <linux/bootmem.h>
13 #include <linux/init.h>
15 #include <linux/mmzone.h>
16 #include <linux/export.h>
17 #include <linux/nodemask.h>
18 #include <linux/cpu.h>
19 #include <linux/notifier.h>
20 #include <linux/memblock.h>
22 #include <linux/pfn.h>
23 #include <linux/cpuset.h>
24 #include <linux/node.h>
25 #include <asm/sparsemem.h>
28 #include <asm/firmware.h>
30 #include <asm/hvcall.h>
31 #include <asm/setup.h>
33 static int numa_enabled = 1;
35 static char *cmdline __initdata;
37 static int numa_debug;
38 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
40 int numa_cpu_lookup_table[NR_CPUS];
41 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
42 struct pglist_data *node_data[MAX_NUMNODES];
44 EXPORT_SYMBOL(numa_cpu_lookup_table);
45 EXPORT_SYMBOL(node_to_cpumask_map);
46 EXPORT_SYMBOL(node_data);
48 static int min_common_depth;
49 static int n_mem_addr_cells, n_mem_size_cells;
50 static int form1_affinity;
52 #define MAX_DISTANCE_REF_POINTS 4
53 static int distance_ref_points_depth;
54 static const unsigned int *distance_ref_points;
55 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
58 * Allocate node_to_cpumask_map based on number of available nodes
59 * Requires node_possible_map to be valid.
61 * Note: cpumask_of_node() is not valid until after this is done.
63 static void __init setup_node_to_cpumask_map(void)
65 unsigned int node, num = 0;
67 /* setup nr_node_ids if not done yet */
68 if (nr_node_ids == MAX_NUMNODES) {
69 for_each_node_mask(node, node_possible_map)
71 nr_node_ids = num + 1;
74 /* allocate the map */
75 for (node = 0; node < nr_node_ids; node++)
76 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
78 /* cpumask_of_node() will now work */
79 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
82 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
85 unsigned long long mem;
87 static unsigned int fake_nid;
88 static unsigned long long curr_boundary;
91 * Modify node id, iff we started creating NUMA nodes
92 * We want to continue from where we left of the last time
97 * In case there are no more arguments to parse, the
98 * node_id should be the same as the last fake node id
99 * (we've handled this above).
104 mem = memparse(p, &p);
108 if (mem < curr_boundary)
113 if ((end_pfn << PAGE_SHIFT) > mem) {
115 * Skip commas and spaces
117 while (*p == ',' || *p == ' ' || *p == '\t')
123 dbg("created new fake_node with id %d\n", fake_nid);
130 * get_node_active_region - Return active region containing pfn
131 * Active range returned is empty if none found.
132 * @pfn: The page to return the region for
133 * @node_ar: Returned set to the active region containing @pfn
135 static void __init get_node_active_region(unsigned long pfn,
136 struct node_active_region *node_ar)
138 unsigned long start_pfn, end_pfn;
141 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
142 if (pfn >= start_pfn && pfn < end_pfn) {
144 node_ar->start_pfn = start_pfn;
145 node_ar->end_pfn = end_pfn;
151 static void map_cpu_to_node(int cpu, int node)
153 numa_cpu_lookup_table[cpu] = node;
155 dbg("adding cpu %d to node %d\n", cpu, node);
157 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
158 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
161 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
162 static void unmap_cpu_from_node(unsigned long cpu)
164 int node = numa_cpu_lookup_table[cpu];
166 dbg("removing cpu %lu from node %d\n", cpu, node);
168 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
169 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
171 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
175 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
177 /* must hold reference to node during call */
178 static const int *of_get_associativity(struct device_node *dev)
180 return of_get_property(dev, "ibm,associativity", NULL);
184 * Returns the property linux,drconf-usable-memory if
185 * it exists (the property exists only in kexec/kdump kernels,
186 * added by kexec-tools)
188 static const u32 *of_get_usable_memory(struct device_node *memory)
192 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
193 if (!prop || len < sizeof(unsigned int))
198 int __node_distance(int a, int b)
201 int distance = LOCAL_DISTANCE;
206 for (i = 0; i < distance_ref_points_depth; i++) {
207 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
210 /* Double the distance for each NUMA level */
217 static void initialize_distance_lookup_table(int nid,
218 const unsigned int *associativity)
225 for (i = 0; i < distance_ref_points_depth; i++) {
226 distance_lookup_table[nid][i] =
227 associativity[distance_ref_points[i]];
231 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
234 static int associativity_to_nid(const unsigned int *associativity)
238 if (min_common_depth == -1)
241 if (associativity[0] >= min_common_depth)
242 nid = associativity[min_common_depth];
244 /* POWER4 LPAR uses 0xffff as invalid node */
245 if (nid == 0xffff || nid >= MAX_NUMNODES)
248 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
249 initialize_distance_lookup_table(nid, associativity);
255 /* Returns the nid associated with the given device tree node,
256 * or -1 if not found.
258 static int of_node_to_nid_single(struct device_node *device)
261 const unsigned int *tmp;
263 tmp = of_get_associativity(device);
265 nid = associativity_to_nid(tmp);
269 /* Walk the device tree upwards, looking for an associativity id */
270 int of_node_to_nid(struct device_node *device)
272 struct device_node *tmp;
277 nid = of_node_to_nid_single(device);
282 device = of_get_parent(tmp);
289 EXPORT_SYMBOL_GPL(of_node_to_nid);
291 static int __init find_min_common_depth(void)
294 struct device_node *chosen;
295 struct device_node *root;
298 if (firmware_has_feature(FW_FEATURE_OPAL))
299 root = of_find_node_by_path("/ibm,opal");
301 root = of_find_node_by_path("/rtas");
303 root = of_find_node_by_path("/");
306 * This property is a set of 32-bit integers, each representing
307 * an index into the ibm,associativity nodes.
309 * With form 0 affinity the first integer is for an SMP configuration
310 * (should be all 0's) and the second is for a normal NUMA
311 * configuration. We have only one level of NUMA.
313 * With form 1 affinity the first integer is the most significant
314 * NUMA boundary and the following are progressively less significant
315 * boundaries. There can be more than one level of NUMA.
317 distance_ref_points = of_get_property(root,
318 "ibm,associativity-reference-points",
319 &distance_ref_points_depth);
321 if (!distance_ref_points) {
322 dbg("NUMA: ibm,associativity-reference-points not found.\n");
326 distance_ref_points_depth /= sizeof(int);
328 #define VEC5_AFFINITY_BYTE 5
329 #define VEC5_AFFINITY 0x80
331 if (firmware_has_feature(FW_FEATURE_OPAL))
334 chosen = of_find_node_by_path("/chosen");
336 vec5 = of_get_property(chosen,
337 "ibm,architecture-vec-5", NULL);
338 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] &
340 dbg("Using form 1 affinity\n");
348 if (form1_affinity) {
349 depth = distance_ref_points[0];
351 if (distance_ref_points_depth < 2) {
352 printk(KERN_WARNING "NUMA: "
353 "short ibm,associativity-reference-points\n");
357 depth = distance_ref_points[1];
361 * Warn and cap if the hardware supports more than
362 * MAX_DISTANCE_REF_POINTS domains.
364 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
365 printk(KERN_WARNING "NUMA: distance array capped at "
366 "%d entries\n", MAX_DISTANCE_REF_POINTS);
367 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
378 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
380 struct device_node *memory = NULL;
382 memory = of_find_node_by_type(memory, "memory");
384 panic("numa.c: No memory nodes found!");
386 *n_addr_cells = of_n_addr_cells(memory);
387 *n_size_cells = of_n_size_cells(memory);
391 static unsigned long read_n_cells(int n, const unsigned int **buf)
393 unsigned long result = 0;
396 result = (result << 32) | **buf;
402 struct of_drconf_cell {
410 #define DRCONF_MEM_ASSIGNED 0x00000008
411 #define DRCONF_MEM_AI_INVALID 0x00000040
412 #define DRCONF_MEM_RESERVED 0x00000080
415 * Read the next memblock list entry from the ibm,dynamic-memory property
416 * and return the information in the provided of_drconf_cell structure.
418 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
422 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
425 drmem->drc_index = cp[0];
426 drmem->reserved = cp[1];
427 drmem->aa_index = cp[2];
428 drmem->flags = cp[3];
434 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
436 * The layout of the ibm,dynamic-memory property is a number N of memblock
437 * list entries followed by N memblock list entries. Each memblock list entry
438 * contains information as laid out in the of_drconf_cell struct above.
440 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
445 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
446 if (!prop || len < sizeof(unsigned int))
451 /* Now that we know the number of entries, revalidate the size
452 * of the property read in to ensure we have everything
454 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
462 * Retrieve and validate the ibm,lmb-size property for drconf memory
463 * from the device tree.
465 static u64 of_get_lmb_size(struct device_node *memory)
470 prop = of_get_property(memory, "ibm,lmb-size", &len);
471 if (!prop || len < sizeof(unsigned int))
474 return read_n_cells(n_mem_size_cells, &prop);
477 struct assoc_arrays {
484 * Retrieve and validate the list of associativity arrays for drconf
485 * memory from the ibm,associativity-lookup-arrays property of the
488 * The layout of the ibm,associativity-lookup-arrays property is a number N
489 * indicating the number of associativity arrays, followed by a number M
490 * indicating the size of each associativity array, followed by a list
491 * of N associativity arrays.
493 static int of_get_assoc_arrays(struct device_node *memory,
494 struct assoc_arrays *aa)
499 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
500 if (!prop || len < 2 * sizeof(unsigned int))
503 aa->n_arrays = *prop++;
504 aa->array_sz = *prop++;
506 /* Now that we know the number of arrays and size of each array,
507 * revalidate the size of the property read in.
509 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
517 * This is like of_node_to_nid_single() for memory represented in the
518 * ibm,dynamic-reconfiguration-memory node.
520 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
521 struct assoc_arrays *aa)
524 int nid = default_nid;
527 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
528 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
529 drmem->aa_index < aa->n_arrays) {
530 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
531 nid = aa->arrays[index];
533 if (nid == 0xffff || nid >= MAX_NUMNODES)
541 * Figure out to which domain a cpu belongs and stick it there.
542 * Return the id of the domain used.
544 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
547 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
554 nid = of_node_to_nid_single(cpu);
556 if (nid < 0 || !node_online(nid))
557 nid = first_online_node;
559 map_cpu_to_node(lcpu, nid);
566 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
567 unsigned long action,
570 unsigned long lcpu = (unsigned long)hcpu;
571 int ret = NOTIFY_DONE;
575 case CPU_UP_PREPARE_FROZEN:
576 numa_setup_cpu(lcpu);
579 #ifdef CONFIG_HOTPLUG_CPU
581 case CPU_DEAD_FROZEN:
582 case CPU_UP_CANCELED:
583 case CPU_UP_CANCELED_FROZEN:
584 unmap_cpu_from_node(lcpu);
593 * Check and possibly modify a memory region to enforce the memory limit.
595 * Returns the size the region should have to enforce the memory limit.
596 * This will either be the original value of size, a truncated value,
597 * or zero. If the returned value of size is 0 the region should be
598 * discarded as it lies wholly above the memory limit.
600 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
604 * We use memblock_end_of_DRAM() in here instead of memory_limit because
605 * we've already adjusted it for the limit and it takes care of
606 * having memory holes below the limit. Also, in the case of
607 * iommu_is_off, memory_limit is not set but is implicitly enforced.
610 if (start + size <= memblock_end_of_DRAM())
613 if (start >= memblock_end_of_DRAM())
616 return memblock_end_of_DRAM() - start;
620 * Reads the counter for a given entry in
621 * linux,drconf-usable-memory property
623 static inline int __init read_usm_ranges(const u32 **usm)
626 * For each lmb in ibm,dynamic-memory a corresponding
627 * entry in linux,drconf-usable-memory property contains
628 * a counter followed by that many (base, size) duple.
629 * read the counter from linux,drconf-usable-memory
631 return read_n_cells(n_mem_size_cells, usm);
635 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
636 * node. This assumes n_mem_{addr,size}_cells have been set.
638 static void __init parse_drconf_memory(struct device_node *memory)
640 const u32 *uninitialized_var(dm), *usm;
641 unsigned int n, rc, ranges, is_kexec_kdump = 0;
642 unsigned long lmb_size, base, size, sz;
644 struct assoc_arrays aa = { .arrays = NULL };
646 n = of_get_drconf_memory(memory, &dm);
650 lmb_size = of_get_lmb_size(memory);
654 rc = of_get_assoc_arrays(memory, &aa);
658 /* check if this is a kexec/kdump kernel */
659 usm = of_get_usable_memory(memory);
663 for (; n != 0; --n) {
664 struct of_drconf_cell drmem;
666 read_drconf_cell(&drmem, &dm);
668 /* skip this block if the reserved bit is set in flags (0x80)
669 or if the block is not assigned to this partition (0x8) */
670 if ((drmem.flags & DRCONF_MEM_RESERVED)
671 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
674 base = drmem.base_addr;
678 if (is_kexec_kdump) {
679 ranges = read_usm_ranges(&usm);
680 if (!ranges) /* there are no (base, size) duple */
684 if (is_kexec_kdump) {
685 base = read_n_cells(n_mem_addr_cells, &usm);
686 size = read_n_cells(n_mem_size_cells, &usm);
688 nid = of_drconf_to_nid_single(&drmem, &aa);
689 fake_numa_create_new_node(
690 ((base + size) >> PAGE_SHIFT),
692 node_set_online(nid);
693 sz = numa_enforce_memory_limit(base, size);
695 memblock_set_node(base, sz, nid);
700 static int __init parse_numa_properties(void)
702 struct device_node *memory;
706 if (numa_enabled == 0) {
707 printk(KERN_WARNING "NUMA disabled by user\n");
711 min_common_depth = find_min_common_depth();
713 if (min_common_depth < 0)
714 return min_common_depth;
716 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
719 * Even though we connect cpus to numa domains later in SMP
720 * init, we need to know the node ids now. This is because
721 * each node to be onlined must have NODE_DATA etc backing it.
723 for_each_present_cpu(i) {
724 struct device_node *cpu;
727 cpu = of_get_cpu_node(i, NULL);
729 nid = of_node_to_nid_single(cpu);
733 * Don't fall back to default_nid yet -- we will plug
734 * cpus into nodes once the memory scan has discovered
739 node_set_online(nid);
742 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
744 for_each_node_by_type(memory, "memory") {
749 const unsigned int *memcell_buf;
752 memcell_buf = of_get_property(memory,
753 "linux,usable-memory", &len);
754 if (!memcell_buf || len <= 0)
755 memcell_buf = of_get_property(memory, "reg", &len);
756 if (!memcell_buf || len <= 0)
760 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
762 /* these are order-sensitive, and modify the buffer pointer */
763 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
764 size = read_n_cells(n_mem_size_cells, &memcell_buf);
767 * Assumption: either all memory nodes or none will
768 * have associativity properties. If none, then
769 * everything goes to default_nid.
771 nid = of_node_to_nid_single(memory);
775 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
776 node_set_online(nid);
778 if (!(size = numa_enforce_memory_limit(start, size))) {
785 memblock_set_node(start, size, nid);
792 * Now do the same thing for each MEMBLOCK listed in the
793 * ibm,dynamic-memory property in the
794 * ibm,dynamic-reconfiguration-memory node.
796 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
798 parse_drconf_memory(memory);
803 static void __init setup_nonnuma(void)
805 unsigned long top_of_ram = memblock_end_of_DRAM();
806 unsigned long total_ram = memblock_phys_mem_size();
807 unsigned long start_pfn, end_pfn;
808 unsigned int nid = 0;
809 struct memblock_region *reg;
811 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
812 top_of_ram, total_ram);
813 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
814 (top_of_ram - total_ram) >> 20);
816 for_each_memblock(memory, reg) {
817 start_pfn = memblock_region_memory_base_pfn(reg);
818 end_pfn = memblock_region_memory_end_pfn(reg);
820 fake_numa_create_new_node(end_pfn, &nid);
821 memblock_set_node(PFN_PHYS(start_pfn),
822 PFN_PHYS(end_pfn - start_pfn), nid);
823 node_set_online(nid);
827 void __init dump_numa_cpu_topology(void)
830 unsigned int cpu, count;
832 if (min_common_depth == -1 || !numa_enabled)
835 for_each_online_node(node) {
836 printk(KERN_DEBUG "Node %d CPUs:", node);
840 * If we used a CPU iterator here we would miss printing
841 * the holes in the cpumap.
843 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
844 if (cpumask_test_cpu(cpu,
845 node_to_cpumask_map[node])) {
851 printk("-%u", cpu - 1);
857 printk("-%u", nr_cpu_ids - 1);
862 static void __init dump_numa_memory_topology(void)
867 if (min_common_depth == -1 || !numa_enabled)
870 for_each_online_node(node) {
873 printk(KERN_DEBUG "Node %d Memory:", node);
877 for (i = 0; i < memblock_end_of_DRAM();
878 i += (1 << SECTION_SIZE_BITS)) {
879 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
897 * Allocate some memory, satisfying the memblock or bootmem allocator where
898 * required. nid is the preferred node and end is the physical address of
899 * the highest address in the node.
901 * Returns the virtual address of the memory.
903 static void __init *careful_zallocation(int nid, unsigned long size,
905 unsigned long end_pfn)
909 unsigned long ret_paddr;
911 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
913 /* retry over all memory */
915 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
918 panic("numa.c: cannot allocate %lu bytes for node %d",
921 ret = __va(ret_paddr);
924 * We initialize the nodes in numeric order: 0, 1, 2...
925 * and hand over control from the MEMBLOCK allocator to the
926 * bootmem allocator. If this function is called for
927 * node 5, then we know that all nodes <5 are using the
928 * bootmem allocator instead of the MEMBLOCK allocator.
930 * So, check the nid from which this allocation came
931 * and double check to see if we need to use bootmem
932 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
933 * since it would be useless.
935 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
937 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
940 dbg("alloc_bootmem %p %lx\n", ret, size);
943 memset(ret, 0, size);
947 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
948 .notifier_call = cpu_numa_callback,
949 .priority = 1 /* Must run before sched domains notifier. */
952 static void __init mark_reserved_regions_for_nid(int nid)
954 struct pglist_data *node = NODE_DATA(nid);
955 struct memblock_region *reg;
957 for_each_memblock(reserved, reg) {
958 unsigned long physbase = reg->base;
959 unsigned long size = reg->size;
960 unsigned long start_pfn = physbase >> PAGE_SHIFT;
961 unsigned long end_pfn = PFN_UP(physbase + size);
962 struct node_active_region node_ar;
963 unsigned long node_end_pfn = node->node_start_pfn +
964 node->node_spanned_pages;
967 * Check to make sure that this memblock.reserved area is
968 * within the bounds of the node that we care about.
969 * Checking the nid of the start and end points is not
970 * sufficient because the reserved area could span the
973 if (end_pfn <= node->node_start_pfn ||
974 start_pfn >= node_end_pfn)
977 get_node_active_region(start_pfn, &node_ar);
978 while (start_pfn < end_pfn &&
979 node_ar.start_pfn < node_ar.end_pfn) {
980 unsigned long reserve_size = size;
982 * if reserved region extends past active region
983 * then trim size to active region
985 if (end_pfn > node_ar.end_pfn)
986 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
989 * Only worry about *this* node, others may not
990 * yet have valid NODE_DATA().
992 if (node_ar.nid == nid) {
993 dbg("reserve_bootmem %lx %lx nid=%d\n",
994 physbase, reserve_size, node_ar.nid);
995 reserve_bootmem_node(NODE_DATA(node_ar.nid),
996 physbase, reserve_size,
1000 * if reserved region is contained in the active region
1003 if (end_pfn <= node_ar.end_pfn)
1007 * reserved region extends past the active region
1008 * get next active region that contains this
1011 start_pfn = node_ar.end_pfn;
1012 physbase = start_pfn << PAGE_SHIFT;
1013 size = size - reserve_size;
1014 get_node_active_region(start_pfn, &node_ar);
1020 void __init do_init_bootmem(void)
1025 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1026 max_pfn = max_low_pfn;
1028 if (parse_numa_properties())
1031 dump_numa_memory_topology();
1033 for_each_online_node(nid) {
1034 unsigned long start_pfn, end_pfn;
1035 void *bootmem_vaddr;
1036 unsigned long bootmap_pages;
1038 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1041 * Allocate the node structure node local if possible
1043 * Be careful moving this around, as it relies on all
1044 * previous nodes' bootmem to be initialized and have
1045 * all reserved areas marked.
1047 NODE_DATA(nid) = careful_zallocation(nid,
1048 sizeof(struct pglist_data),
1049 SMP_CACHE_BYTES, end_pfn);
1051 dbg("node %d\n", nid);
1052 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1054 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1055 NODE_DATA(nid)->node_start_pfn = start_pfn;
1056 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1058 if (NODE_DATA(nid)->node_spanned_pages == 0)
1061 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1062 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1064 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1065 bootmem_vaddr = careful_zallocation(nid,
1066 bootmap_pages << PAGE_SHIFT,
1067 PAGE_SIZE, end_pfn);
1069 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1071 init_bootmem_node(NODE_DATA(nid),
1072 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1073 start_pfn, end_pfn);
1075 free_bootmem_with_active_regions(nid, end_pfn);
1077 * Be very careful about moving this around. Future
1078 * calls to careful_zallocation() depend on this getting
1081 mark_reserved_regions_for_nid(nid);
1082 sparse_memory_present_with_active_regions(nid);
1085 init_bootmem_done = 1;
1088 * Now bootmem is initialised we can create the node to cpumask
1089 * lookup tables and setup the cpu callback to populate them.
1091 setup_node_to_cpumask_map();
1093 register_cpu_notifier(&ppc64_numa_nb);
1094 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1095 (void *)(unsigned long)boot_cpuid);
1098 void __init paging_init(void)
1100 unsigned long max_zone_pfns[MAX_NR_ZONES];
1101 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1102 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1103 free_area_init_nodes(max_zone_pfns);
1106 static int __init early_numa(char *p)
1111 if (strstr(p, "off"))
1114 if (strstr(p, "debug"))
1117 p = strstr(p, "fake=");
1119 cmdline = p + strlen("fake=");
1123 early_param("numa", early_numa);
1125 #ifdef CONFIG_MEMORY_HOTPLUG
1127 * Find the node associated with a hot added memory section for
1128 * memory represented in the device tree by the property
1129 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1131 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1132 unsigned long scn_addr)
1135 unsigned int drconf_cell_cnt, rc;
1136 unsigned long lmb_size;
1137 struct assoc_arrays aa;
1140 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1141 if (!drconf_cell_cnt)
1144 lmb_size = of_get_lmb_size(memory);
1148 rc = of_get_assoc_arrays(memory, &aa);
1152 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1153 struct of_drconf_cell drmem;
1155 read_drconf_cell(&drmem, &dm);
1157 /* skip this block if it is reserved or not assigned to
1159 if ((drmem.flags & DRCONF_MEM_RESERVED)
1160 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1163 if ((scn_addr < drmem.base_addr)
1164 || (scn_addr >= (drmem.base_addr + lmb_size)))
1167 nid = of_drconf_to_nid_single(&drmem, &aa);
1175 * Find the node associated with a hot added memory section for memory
1176 * represented in the device tree as a node (i.e. memory@XXXX) for
1179 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1181 struct device_node *memory;
1184 for_each_node_by_type(memory, "memory") {
1185 unsigned long start, size;
1187 const unsigned int *memcell_buf;
1190 memcell_buf = of_get_property(memory, "reg", &len);
1191 if (!memcell_buf || len <= 0)
1194 /* ranges in cell */
1195 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1198 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1199 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1201 if ((scn_addr < start) || (scn_addr >= (start + size)))
1204 nid = of_node_to_nid_single(memory);
1212 of_node_put(memory);
1218 * Find the node associated with a hot added memory section. Section
1219 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1220 * sections are fully contained within a single MEMBLOCK.
1222 int hot_add_scn_to_nid(unsigned long scn_addr)
1224 struct device_node *memory = NULL;
1227 if (!numa_enabled || (min_common_depth < 0))
1228 return first_online_node;
1230 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1232 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1233 of_node_put(memory);
1235 nid = hot_add_node_scn_to_nid(scn_addr);
1238 if (nid < 0 || !node_online(nid))
1239 nid = first_online_node;
1241 if (NODE_DATA(nid)->node_spanned_pages)
1244 for_each_online_node(nid) {
1245 if (NODE_DATA(nid)->node_spanned_pages) {
1255 static u64 hot_add_drconf_memory_max(void)
1257 struct device_node *memory = NULL;
1258 unsigned int drconf_cell_cnt = 0;
1262 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1264 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1265 lmb_size = of_get_lmb_size(memory);
1266 of_node_put(memory);
1268 return lmb_size * drconf_cell_cnt;
1272 * memory_hotplug_max - return max address of memory that may be added
1274 * This is currently only used on systems that support drconfig memory
1277 u64 memory_hotplug_max(void)
1279 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1281 #endif /* CONFIG_MEMORY_HOTPLUG */
1283 /* Virtual Processor Home Node (VPHN) support */
1284 #ifdef CONFIG_PPC_SPLPAR
1285 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1286 static cpumask_t cpu_associativity_changes_mask;
1287 static int vphn_enabled;
1288 static void set_topology_timer(void);
1291 * Store the current values of the associativity change counters in the
1294 static void setup_cpu_associativity_change_counters(void)
1298 /* The VPHN feature supports a maximum of 8 reference points */
1299 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1301 for_each_possible_cpu(cpu) {
1303 u8 *counts = vphn_cpu_change_counts[cpu];
1304 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1306 for (i = 0; i < distance_ref_points_depth; i++)
1307 counts[i] = hypervisor_counts[i];
1312 * The hypervisor maintains a set of 8 associativity change counters in
1313 * the VPA of each cpu that correspond to the associativity levels in the
1314 * ibm,associativity-reference-points property. When an associativity
1315 * level changes, the corresponding counter is incremented.
1317 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1318 * node associativity levels have changed.
1320 * Returns the number of cpus with unhandled associativity changes.
1322 static int update_cpu_associativity_changes_mask(void)
1324 int cpu, nr_cpus = 0;
1325 cpumask_t *changes = &cpu_associativity_changes_mask;
1327 cpumask_clear(changes);
1329 for_each_possible_cpu(cpu) {
1331 u8 *counts = vphn_cpu_change_counts[cpu];
1332 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1334 for (i = 0; i < distance_ref_points_depth; i++) {
1335 if (hypervisor_counts[i] != counts[i]) {
1336 counts[i] = hypervisor_counts[i];
1341 cpumask_set_cpu(cpu, changes);
1350 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1351 * the complete property we have to add the length in the first cell.
1353 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1356 * Convert the associativity domain numbers returned from the hypervisor
1357 * to the sequence they would appear in the ibm,associativity property.
1359 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1361 int i, nr_assoc_doms = 0;
1362 const u16 *field = (const u16*) packed;
1364 #define VPHN_FIELD_UNUSED (0xffff)
1365 #define VPHN_FIELD_MSB (0x8000)
1366 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1368 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1369 if (*field == VPHN_FIELD_UNUSED) {
1370 /* All significant fields processed, and remaining
1371 * fields contain the reserved value of all 1's.
1374 unpacked[i] = *((u32*)field);
1376 } else if (*field & VPHN_FIELD_MSB) {
1377 /* Data is in the lower 15 bits of this field */
1378 unpacked[i] = *field & VPHN_FIELD_MASK;
1382 /* Data is in the lower 15 bits of this field
1383 * concatenated with the next 16 bit field
1385 unpacked[i] = *((u32*)field);
1391 /* The first cell contains the length of the property */
1392 unpacked[0] = nr_assoc_doms;
1394 return nr_assoc_doms;
1398 * Retrieve the new associativity information for a virtual processor's
1401 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1404 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1406 int hwcpu = get_hard_smp_processor_id(cpu);
1408 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1409 vphn_unpack_associativity(retbuf, associativity);
1414 static long vphn_get_associativity(unsigned long cpu,
1415 unsigned int *associativity)
1419 rc = hcall_vphn(cpu, associativity);
1424 "VPHN is not supported. Disabling polling...\n");
1425 stop_topology_update();
1429 "hcall_vphn() experienced a hardware fault "
1430 "preventing VPHN. Disabling polling...\n");
1431 stop_topology_update();
1438 * Update the node maps and sysfs entries for each cpu whose home node
1439 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1441 int arch_update_cpu_topology(void)
1443 int cpu, nid, old_nid, changed = 0;
1444 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1447 for_each_cpu(cpu,&cpu_associativity_changes_mask) {
1448 vphn_get_associativity(cpu, associativity);
1449 nid = associativity_to_nid(associativity);
1451 if (nid < 0 || !node_online(nid))
1452 nid = first_online_node;
1454 old_nid = numa_cpu_lookup_table[cpu];
1456 /* Disable hotplug while we update the cpu
1460 unregister_cpu_under_node(cpu, old_nid);
1461 unmap_cpu_from_node(cpu);
1462 map_cpu_to_node(cpu, nid);
1463 register_cpu_under_node(cpu, nid);
1466 dev = get_cpu_device(cpu);
1468 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1475 static void topology_work_fn(struct work_struct *work)
1477 rebuild_sched_domains();
1479 static DECLARE_WORK(topology_work, topology_work_fn);
1481 void topology_schedule_update(void)
1483 schedule_work(&topology_work);
1486 static void topology_timer_fn(unsigned long ignored)
1490 if (update_cpu_associativity_changes_mask() > 0)
1491 topology_schedule_update();
1492 set_topology_timer();
1494 static struct timer_list topology_timer =
1495 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1497 static void set_topology_timer(void)
1499 topology_timer.data = 0;
1500 topology_timer.expires = jiffies + 60 * HZ;
1501 add_timer(&topology_timer);
1505 * Start polling for VPHN associativity changes.
1507 int start_topology_update(void)
1511 /* Disabled until races with load balancing are fixed */
1512 if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1513 get_lppaca()->shared_proc) {
1515 setup_cpu_associativity_change_counters();
1516 init_timer_deferrable(&topology_timer);
1517 set_topology_timer();
1523 __initcall(start_topology_update);
1526 * Disable polling for VPHN associativity changes.
1528 int stop_topology_update(void)
1531 return del_timer_sync(&topology_timer);
1533 #endif /* CONFIG_PPC_SPLPAR */