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>
27 #include <asm/system.h>
29 #include <asm/firmware.h>
31 #include <asm/hvcall.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: node_to_cpumask() 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_active_region_work_fn - A helper function for get_node_active_region
131 * Returns datax set to the start_pfn and end_pfn if they contain
132 * the initial value of datax->start_pfn between them
133 * @start_pfn: start page(inclusive) of region to check
134 * @end_pfn: end page(exclusive) of region to check
135 * @datax: comes in with ->start_pfn set to value to search for and
136 * goes out with active range if it contains it
137 * Returns 1 if search value is in range else 0
139 static int __init get_active_region_work_fn(unsigned long start_pfn,
140 unsigned long end_pfn, void *datax)
142 struct node_active_region *data;
143 data = (struct node_active_region *)datax;
145 if (start_pfn <= data->start_pfn && end_pfn > data->start_pfn) {
146 data->start_pfn = start_pfn;
147 data->end_pfn = end_pfn;
155 * get_node_active_region - Return active region containing start_pfn
156 * Active range returned is empty if none found.
157 * @start_pfn: The page to return the region for.
158 * @node_ar: Returned set to the active region containing start_pfn
160 static void __init get_node_active_region(unsigned long start_pfn,
161 struct node_active_region *node_ar)
163 int nid = early_pfn_to_nid(start_pfn);
166 node_ar->start_pfn = start_pfn;
167 node_ar->end_pfn = start_pfn;
168 work_with_active_regions(nid, get_active_region_work_fn, node_ar);
171 static void map_cpu_to_node(int cpu, int node)
173 numa_cpu_lookup_table[cpu] = node;
175 dbg("adding cpu %d to node %d\n", cpu, node);
177 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
178 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
181 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
182 static void unmap_cpu_from_node(unsigned long cpu)
184 int node = numa_cpu_lookup_table[cpu];
186 dbg("removing cpu %lu from node %d\n", cpu, node);
188 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
189 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
191 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
195 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
197 /* must hold reference to node during call */
198 static const int *of_get_associativity(struct device_node *dev)
200 return of_get_property(dev, "ibm,associativity", NULL);
204 * Returns the property linux,drconf-usable-memory if
205 * it exists (the property exists only in kexec/kdump kernels,
206 * added by kexec-tools)
208 static const u32 *of_get_usable_memory(struct device_node *memory)
212 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
213 if (!prop || len < sizeof(unsigned int))
218 int __node_distance(int a, int b)
221 int distance = LOCAL_DISTANCE;
226 for (i = 0; i < distance_ref_points_depth; i++) {
227 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
230 /* Double the distance for each NUMA level */
237 static void initialize_distance_lookup_table(int nid,
238 const unsigned int *associativity)
245 for (i = 0; i < distance_ref_points_depth; i++) {
246 distance_lookup_table[nid][i] =
247 associativity[distance_ref_points[i]];
251 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
254 static int associativity_to_nid(const unsigned int *associativity)
258 if (min_common_depth == -1)
261 if (associativity[0] >= min_common_depth)
262 nid = associativity[min_common_depth];
264 /* POWER4 LPAR uses 0xffff as invalid node */
265 if (nid == 0xffff || nid >= MAX_NUMNODES)
268 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
269 initialize_distance_lookup_table(nid, associativity);
275 /* Returns the nid associated with the given device tree node,
276 * or -1 if not found.
278 static int of_node_to_nid_single(struct device_node *device)
281 const unsigned int *tmp;
283 tmp = of_get_associativity(device);
285 nid = associativity_to_nid(tmp);
289 /* Walk the device tree upwards, looking for an associativity id */
290 int of_node_to_nid(struct device_node *device)
292 struct device_node *tmp;
297 nid = of_node_to_nid_single(device);
302 device = of_get_parent(tmp);
309 EXPORT_SYMBOL_GPL(of_node_to_nid);
311 static int __init find_min_common_depth(void)
314 struct device_node *chosen;
315 struct device_node *root;
318 if (firmware_has_feature(FW_FEATURE_OPAL))
319 root = of_find_node_by_path("/ibm,opal");
321 root = of_find_node_by_path("/rtas");
323 root = of_find_node_by_path("/");
326 * This property is a set of 32-bit integers, each representing
327 * an index into the ibm,associativity nodes.
329 * With form 0 affinity the first integer is for an SMP configuration
330 * (should be all 0's) and the second is for a normal NUMA
331 * configuration. We have only one level of NUMA.
333 * With form 1 affinity the first integer is the most significant
334 * NUMA boundary and the following are progressively less significant
335 * boundaries. There can be more than one level of NUMA.
337 distance_ref_points = of_get_property(root,
338 "ibm,associativity-reference-points",
339 &distance_ref_points_depth);
341 if (!distance_ref_points) {
342 dbg("NUMA: ibm,associativity-reference-points not found.\n");
346 distance_ref_points_depth /= sizeof(int);
348 #define VEC5_AFFINITY_BYTE 5
349 #define VEC5_AFFINITY 0x80
351 if (firmware_has_feature(FW_FEATURE_OPAL))
354 chosen = of_find_node_by_path("/chosen");
356 vec5 = of_get_property(chosen,
357 "ibm,architecture-vec-5", NULL);
358 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] &
360 dbg("Using form 1 affinity\n");
366 if (form1_affinity) {
367 depth = distance_ref_points[0];
369 if (distance_ref_points_depth < 2) {
370 printk(KERN_WARNING "NUMA: "
371 "short ibm,associativity-reference-points\n");
375 depth = distance_ref_points[1];
379 * Warn and cap if the hardware supports more than
380 * MAX_DISTANCE_REF_POINTS domains.
382 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
383 printk(KERN_WARNING "NUMA: distance array capped at "
384 "%d entries\n", MAX_DISTANCE_REF_POINTS);
385 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
396 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
398 struct device_node *memory = NULL;
400 memory = of_find_node_by_type(memory, "memory");
402 panic("numa.c: No memory nodes found!");
404 *n_addr_cells = of_n_addr_cells(memory);
405 *n_size_cells = of_n_size_cells(memory);
409 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
411 unsigned long result = 0;
414 result = (result << 32) | **buf;
420 struct of_drconf_cell {
428 #define DRCONF_MEM_ASSIGNED 0x00000008
429 #define DRCONF_MEM_AI_INVALID 0x00000040
430 #define DRCONF_MEM_RESERVED 0x00000080
433 * Read the next memblock list entry from the ibm,dynamic-memory property
434 * and return the information in the provided of_drconf_cell structure.
436 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
440 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
443 drmem->drc_index = cp[0];
444 drmem->reserved = cp[1];
445 drmem->aa_index = cp[2];
446 drmem->flags = cp[3];
452 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
454 * The layout of the ibm,dynamic-memory property is a number N of memblock
455 * list entries followed by N memblock list entries. Each memblock list entry
456 * contains information as laid out in the of_drconf_cell struct above.
458 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
463 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
464 if (!prop || len < sizeof(unsigned int))
469 /* Now that we know the number of entries, revalidate the size
470 * of the property read in to ensure we have everything
472 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
480 * Retrieve and validate the ibm,lmb-size property for drconf memory
481 * from the device tree.
483 static u64 of_get_lmb_size(struct device_node *memory)
488 prop = of_get_property(memory, "ibm,lmb-size", &len);
489 if (!prop || len < sizeof(unsigned int))
492 return read_n_cells(n_mem_size_cells, &prop);
495 struct assoc_arrays {
502 * Retrieve and validate the list of associativity arrays for drconf
503 * memory from the ibm,associativity-lookup-arrays property of the
506 * The layout of the ibm,associativity-lookup-arrays property is a number N
507 * indicating the number of associativity arrays, followed by a number M
508 * indicating the size of each associativity array, followed by a list
509 * of N associativity arrays.
511 static int of_get_assoc_arrays(struct device_node *memory,
512 struct assoc_arrays *aa)
517 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
518 if (!prop || len < 2 * sizeof(unsigned int))
521 aa->n_arrays = *prop++;
522 aa->array_sz = *prop++;
524 /* Now that we know the number of arrrays and size of each array,
525 * revalidate the size of the property read in.
527 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
535 * This is like of_node_to_nid_single() for memory represented in the
536 * ibm,dynamic-reconfiguration-memory node.
538 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
539 struct assoc_arrays *aa)
542 int nid = default_nid;
545 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
546 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
547 drmem->aa_index < aa->n_arrays) {
548 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
549 nid = aa->arrays[index];
551 if (nid == 0xffff || nid >= MAX_NUMNODES)
559 * Figure out to which domain a cpu belongs and stick it there.
560 * Return the id of the domain used.
562 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
565 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
572 nid = of_node_to_nid_single(cpu);
574 if (nid < 0 || !node_online(nid))
575 nid = first_online_node;
577 map_cpu_to_node(lcpu, nid);
584 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
585 unsigned long action,
588 unsigned long lcpu = (unsigned long)hcpu;
589 int ret = NOTIFY_DONE;
593 case CPU_UP_PREPARE_FROZEN:
594 numa_setup_cpu(lcpu);
597 #ifdef CONFIG_HOTPLUG_CPU
599 case CPU_DEAD_FROZEN:
600 case CPU_UP_CANCELED:
601 case CPU_UP_CANCELED_FROZEN:
602 unmap_cpu_from_node(lcpu);
611 * Check and possibly modify a memory region to enforce the memory limit.
613 * Returns the size the region should have to enforce the memory limit.
614 * This will either be the original value of size, a truncated value,
615 * or zero. If the returned value of size is 0 the region should be
616 * discarded as it lies wholly above the memory limit.
618 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
622 * We use memblock_end_of_DRAM() in here instead of memory_limit because
623 * we've already adjusted it for the limit and it takes care of
624 * having memory holes below the limit. Also, in the case of
625 * iommu_is_off, memory_limit is not set but is implicitly enforced.
628 if (start + size <= memblock_end_of_DRAM())
631 if (start >= memblock_end_of_DRAM())
634 return memblock_end_of_DRAM() - start;
638 * Reads the counter for a given entry in
639 * linux,drconf-usable-memory property
641 static inline int __init read_usm_ranges(const u32 **usm)
644 * For each lmb in ibm,dynamic-memory a corresponding
645 * entry in linux,drconf-usable-memory property contains
646 * a counter followed by that many (base, size) duple.
647 * read the counter from linux,drconf-usable-memory
649 return read_n_cells(n_mem_size_cells, usm);
653 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
654 * node. This assumes n_mem_{addr,size}_cells have been set.
656 static void __init parse_drconf_memory(struct device_node *memory)
659 unsigned int n, rc, ranges, is_kexec_kdump = 0;
660 unsigned long lmb_size, base, size, sz;
662 struct assoc_arrays aa;
664 n = of_get_drconf_memory(memory, &dm);
668 lmb_size = of_get_lmb_size(memory);
672 rc = of_get_assoc_arrays(memory, &aa);
676 /* check if this is a kexec/kdump kernel */
677 usm = of_get_usable_memory(memory);
681 for (; n != 0; --n) {
682 struct of_drconf_cell drmem;
684 read_drconf_cell(&drmem, &dm);
686 /* skip this block if the reserved bit is set in flags (0x80)
687 or if the block is not assigned to this partition (0x8) */
688 if ((drmem.flags & DRCONF_MEM_RESERVED)
689 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
692 base = drmem.base_addr;
696 if (is_kexec_kdump) {
697 ranges = read_usm_ranges(&usm);
698 if (!ranges) /* there are no (base, size) duple */
702 if (is_kexec_kdump) {
703 base = read_n_cells(n_mem_addr_cells, &usm);
704 size = read_n_cells(n_mem_size_cells, &usm);
706 nid = of_drconf_to_nid_single(&drmem, &aa);
707 fake_numa_create_new_node(
708 ((base + size) >> PAGE_SHIFT),
710 node_set_online(nid);
711 sz = numa_enforce_memory_limit(base, size);
713 add_active_range(nid, base >> PAGE_SHIFT,
715 + (sz >> PAGE_SHIFT));
720 static int __init parse_numa_properties(void)
722 struct device_node *memory;
726 if (numa_enabled == 0) {
727 printk(KERN_WARNING "NUMA disabled by user\n");
731 min_common_depth = find_min_common_depth();
733 if (min_common_depth < 0)
734 return min_common_depth;
736 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
739 * Even though we connect cpus to numa domains later in SMP
740 * init, we need to know the node ids now. This is because
741 * each node to be onlined must have NODE_DATA etc backing it.
743 for_each_present_cpu(i) {
744 struct device_node *cpu;
747 cpu = of_get_cpu_node(i, NULL);
749 nid = of_node_to_nid_single(cpu);
753 * Don't fall back to default_nid yet -- we will plug
754 * cpus into nodes once the memory scan has discovered
759 node_set_online(nid);
762 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
764 for_each_node_by_type(memory, "memory") {
769 const unsigned int *memcell_buf;
772 memcell_buf = of_get_property(memory,
773 "linux,usable-memory", &len);
774 if (!memcell_buf || len <= 0)
775 memcell_buf = of_get_property(memory, "reg", &len);
776 if (!memcell_buf || len <= 0)
780 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
782 /* these are order-sensitive, and modify the buffer pointer */
783 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
784 size = read_n_cells(n_mem_size_cells, &memcell_buf);
787 * Assumption: either all memory nodes or none will
788 * have associativity properties. If none, then
789 * everything goes to default_nid.
791 nid = of_node_to_nid_single(memory);
795 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
796 node_set_online(nid);
798 if (!(size = numa_enforce_memory_limit(start, size))) {
805 add_active_range(nid, start >> PAGE_SHIFT,
806 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
813 * Now do the same thing for each MEMBLOCK listed in the
814 * ibm,dynamic-memory property in the
815 * ibm,dynamic-reconfiguration-memory node.
817 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
819 parse_drconf_memory(memory);
824 static void __init setup_nonnuma(void)
826 unsigned long top_of_ram = memblock_end_of_DRAM();
827 unsigned long total_ram = memblock_phys_mem_size();
828 unsigned long start_pfn, end_pfn;
829 unsigned int nid = 0;
830 struct memblock_region *reg;
832 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
833 top_of_ram, total_ram);
834 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
835 (top_of_ram - total_ram) >> 20);
837 for_each_memblock(memory, reg) {
838 start_pfn = memblock_region_memory_base_pfn(reg);
839 end_pfn = memblock_region_memory_end_pfn(reg);
841 fake_numa_create_new_node(end_pfn, &nid);
842 add_active_range(nid, start_pfn, end_pfn);
843 node_set_online(nid);
847 void __init dump_numa_cpu_topology(void)
850 unsigned int cpu, count;
852 if (min_common_depth == -1 || !numa_enabled)
855 for_each_online_node(node) {
856 printk(KERN_DEBUG "Node %d CPUs:", node);
860 * If we used a CPU iterator here we would miss printing
861 * the holes in the cpumap.
863 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
864 if (cpumask_test_cpu(cpu,
865 node_to_cpumask_map[node])) {
871 printk("-%u", cpu - 1);
877 printk("-%u", nr_cpu_ids - 1);
882 static void __init dump_numa_memory_topology(void)
887 if (min_common_depth == -1 || !numa_enabled)
890 for_each_online_node(node) {
893 printk(KERN_DEBUG "Node %d Memory:", node);
897 for (i = 0; i < memblock_end_of_DRAM();
898 i += (1 << SECTION_SIZE_BITS)) {
899 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
917 * Allocate some memory, satisfying the memblock or bootmem allocator where
918 * required. nid is the preferred node and end is the physical address of
919 * the highest address in the node.
921 * Returns the virtual address of the memory.
923 static void __init *careful_zallocation(int nid, unsigned long size,
925 unsigned long end_pfn)
929 unsigned long ret_paddr;
931 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
933 /* retry over all memory */
935 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
938 panic("numa.c: cannot allocate %lu bytes for node %d",
941 ret = __va(ret_paddr);
944 * We initialize the nodes in numeric order: 0, 1, 2...
945 * and hand over control from the MEMBLOCK allocator to the
946 * bootmem allocator. If this function is called for
947 * node 5, then we know that all nodes <5 are using the
948 * bootmem allocator instead of the MEMBLOCK allocator.
950 * So, check the nid from which this allocation came
951 * and double check to see if we need to use bootmem
952 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
953 * since it would be useless.
955 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
957 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
960 dbg("alloc_bootmem %p %lx\n", ret, size);
963 memset(ret, 0, size);
967 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
968 .notifier_call = cpu_numa_callback,
969 .priority = 1 /* Must run before sched domains notifier. */
972 static void mark_reserved_regions_for_nid(int nid)
974 struct pglist_data *node = NODE_DATA(nid);
975 struct memblock_region *reg;
977 for_each_memblock(reserved, reg) {
978 unsigned long physbase = reg->base;
979 unsigned long size = reg->size;
980 unsigned long start_pfn = physbase >> PAGE_SHIFT;
981 unsigned long end_pfn = PFN_UP(physbase + size);
982 struct node_active_region node_ar;
983 unsigned long node_end_pfn = node->node_start_pfn +
984 node->node_spanned_pages;
987 * Check to make sure that this memblock.reserved area is
988 * within the bounds of the node that we care about.
989 * Checking the nid of the start and end points is not
990 * sufficient because the reserved area could span the
993 if (end_pfn <= node->node_start_pfn ||
994 start_pfn >= node_end_pfn)
997 get_node_active_region(start_pfn, &node_ar);
998 while (start_pfn < end_pfn &&
999 node_ar.start_pfn < node_ar.end_pfn) {
1000 unsigned long reserve_size = size;
1002 * if reserved region extends past active region
1003 * then trim size to active region
1005 if (end_pfn > node_ar.end_pfn)
1006 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1009 * Only worry about *this* node, others may not
1010 * yet have valid NODE_DATA().
1012 if (node_ar.nid == nid) {
1013 dbg("reserve_bootmem %lx %lx nid=%d\n",
1014 physbase, reserve_size, node_ar.nid);
1015 reserve_bootmem_node(NODE_DATA(node_ar.nid),
1016 physbase, reserve_size,
1020 * if reserved region is contained in the active region
1023 if (end_pfn <= node_ar.end_pfn)
1027 * reserved region extends past the active region
1028 * get next active region that contains this
1031 start_pfn = node_ar.end_pfn;
1032 physbase = start_pfn << PAGE_SHIFT;
1033 size = size - reserve_size;
1034 get_node_active_region(start_pfn, &node_ar);
1040 void __init do_init_bootmem(void)
1045 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1046 max_pfn = max_low_pfn;
1048 if (parse_numa_properties())
1051 dump_numa_memory_topology();
1053 for_each_online_node(nid) {
1054 unsigned long start_pfn, end_pfn;
1055 void *bootmem_vaddr;
1056 unsigned long bootmap_pages;
1058 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1061 * Allocate the node structure node local if possible
1063 * Be careful moving this around, as it relies on all
1064 * previous nodes' bootmem to be initialized and have
1065 * all reserved areas marked.
1067 NODE_DATA(nid) = careful_zallocation(nid,
1068 sizeof(struct pglist_data),
1069 SMP_CACHE_BYTES, end_pfn);
1071 dbg("node %d\n", nid);
1072 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1074 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1075 NODE_DATA(nid)->node_start_pfn = start_pfn;
1076 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1078 if (NODE_DATA(nid)->node_spanned_pages == 0)
1081 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1082 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1084 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1085 bootmem_vaddr = careful_zallocation(nid,
1086 bootmap_pages << PAGE_SHIFT,
1087 PAGE_SIZE, end_pfn);
1089 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1091 init_bootmem_node(NODE_DATA(nid),
1092 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1093 start_pfn, end_pfn);
1095 free_bootmem_with_active_regions(nid, end_pfn);
1097 * Be very careful about moving this around. Future
1098 * calls to careful_zallocation() depend on this getting
1101 mark_reserved_regions_for_nid(nid);
1102 sparse_memory_present_with_active_regions(nid);
1105 init_bootmem_done = 1;
1108 * Now bootmem is initialised we can create the node to cpumask
1109 * lookup tables and setup the cpu callback to populate them.
1111 setup_node_to_cpumask_map();
1113 register_cpu_notifier(&ppc64_numa_nb);
1114 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1115 (void *)(unsigned long)boot_cpuid);
1118 void __init paging_init(void)
1120 unsigned long max_zone_pfns[MAX_NR_ZONES];
1121 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1122 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1123 free_area_init_nodes(max_zone_pfns);
1126 static int __init early_numa(char *p)
1131 if (strstr(p, "off"))
1134 if (strstr(p, "debug"))
1137 p = strstr(p, "fake=");
1139 cmdline = p + strlen("fake=");
1143 early_param("numa", early_numa);
1145 #ifdef CONFIG_MEMORY_HOTPLUG
1147 * Find the node associated with a hot added memory section for
1148 * memory represented in the device tree by the property
1149 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1151 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1152 unsigned long scn_addr)
1155 unsigned int drconf_cell_cnt, rc;
1156 unsigned long lmb_size;
1157 struct assoc_arrays aa;
1160 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1161 if (!drconf_cell_cnt)
1164 lmb_size = of_get_lmb_size(memory);
1168 rc = of_get_assoc_arrays(memory, &aa);
1172 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1173 struct of_drconf_cell drmem;
1175 read_drconf_cell(&drmem, &dm);
1177 /* skip this block if it is reserved or not assigned to
1179 if ((drmem.flags & DRCONF_MEM_RESERVED)
1180 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1183 if ((scn_addr < drmem.base_addr)
1184 || (scn_addr >= (drmem.base_addr + lmb_size)))
1187 nid = of_drconf_to_nid_single(&drmem, &aa);
1195 * Find the node associated with a hot added memory section for memory
1196 * represented in the device tree as a node (i.e. memory@XXXX) for
1199 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1201 struct device_node *memory;
1204 for_each_node_by_type(memory, "memory") {
1205 unsigned long start, size;
1207 const unsigned int *memcell_buf;
1210 memcell_buf = of_get_property(memory, "reg", &len);
1211 if (!memcell_buf || len <= 0)
1214 /* ranges in cell */
1215 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1218 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1219 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1221 if ((scn_addr < start) || (scn_addr >= (start + size)))
1224 nid = of_node_to_nid_single(memory);
1232 of_node_put(memory);
1238 * Find the node associated with a hot added memory section. Section
1239 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1240 * sections are fully contained within a single MEMBLOCK.
1242 int hot_add_scn_to_nid(unsigned long scn_addr)
1244 struct device_node *memory = NULL;
1247 if (!numa_enabled || (min_common_depth < 0))
1248 return first_online_node;
1250 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1252 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1253 of_node_put(memory);
1255 nid = hot_add_node_scn_to_nid(scn_addr);
1258 if (nid < 0 || !node_online(nid))
1259 nid = first_online_node;
1261 if (NODE_DATA(nid)->node_spanned_pages)
1264 for_each_online_node(nid) {
1265 if (NODE_DATA(nid)->node_spanned_pages) {
1275 static u64 hot_add_drconf_memory_max(void)
1277 struct device_node *memory = NULL;
1278 unsigned int drconf_cell_cnt = 0;
1282 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1284 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1285 lmb_size = of_get_lmb_size(memory);
1286 of_node_put(memory);
1288 return lmb_size * drconf_cell_cnt;
1292 * memory_hotplug_max - return max address of memory that may be added
1294 * This is currently only used on systems that support drconfig memory
1297 u64 memory_hotplug_max(void)
1299 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1301 #endif /* CONFIG_MEMORY_HOTPLUG */
1303 /* Virtual Processor Home Node (VPHN) support */
1304 #ifdef CONFIG_PPC_SPLPAR
1305 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1306 static cpumask_t cpu_associativity_changes_mask;
1307 static int vphn_enabled;
1308 static void set_topology_timer(void);
1311 * Store the current values of the associativity change counters in the
1314 static void setup_cpu_associativity_change_counters(void)
1318 /* The VPHN feature supports a maximum of 8 reference points */
1319 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1321 for_each_possible_cpu(cpu) {
1323 u8 *counts = vphn_cpu_change_counts[cpu];
1324 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1326 for (i = 0; i < distance_ref_points_depth; i++)
1327 counts[i] = hypervisor_counts[i];
1332 * The hypervisor maintains a set of 8 associativity change counters in
1333 * the VPA of each cpu that correspond to the associativity levels in the
1334 * ibm,associativity-reference-points property. When an associativity
1335 * level changes, the corresponding counter is incremented.
1337 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1338 * node associativity levels have changed.
1340 * Returns the number of cpus with unhandled associativity changes.
1342 static int update_cpu_associativity_changes_mask(void)
1344 int cpu, nr_cpus = 0;
1345 cpumask_t *changes = &cpu_associativity_changes_mask;
1347 cpumask_clear(changes);
1349 for_each_possible_cpu(cpu) {
1351 u8 *counts = vphn_cpu_change_counts[cpu];
1352 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1354 for (i = 0; i < distance_ref_points_depth; i++) {
1355 if (hypervisor_counts[i] != counts[i]) {
1356 counts[i] = hypervisor_counts[i];
1361 cpumask_set_cpu(cpu, changes);
1370 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1371 * the complete property we have to add the length in the first cell.
1373 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1376 * Convert the associativity domain numbers returned from the hypervisor
1377 * to the sequence they would appear in the ibm,associativity property.
1379 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1381 int i, nr_assoc_doms = 0;
1382 const u16 *field = (const u16*) packed;
1384 #define VPHN_FIELD_UNUSED (0xffff)
1385 #define VPHN_FIELD_MSB (0x8000)
1386 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1388 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1389 if (*field == VPHN_FIELD_UNUSED) {
1390 /* All significant fields processed, and remaining
1391 * fields contain the reserved value of all 1's.
1394 unpacked[i] = *((u32*)field);
1396 } else if (*field & VPHN_FIELD_MSB) {
1397 /* Data is in the lower 15 bits of this field */
1398 unpacked[i] = *field & VPHN_FIELD_MASK;
1402 /* Data is in the lower 15 bits of this field
1403 * concatenated with the next 16 bit field
1405 unpacked[i] = *((u32*)field);
1411 /* The first cell contains the length of the property */
1412 unpacked[0] = nr_assoc_doms;
1414 return nr_assoc_doms;
1418 * Retrieve the new associativity information for a virtual processor's
1421 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1424 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1426 int hwcpu = get_hard_smp_processor_id(cpu);
1428 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1429 vphn_unpack_associativity(retbuf, associativity);
1434 static long vphn_get_associativity(unsigned long cpu,
1435 unsigned int *associativity)
1439 rc = hcall_vphn(cpu, associativity);
1444 "VPHN is not supported. Disabling polling...\n");
1445 stop_topology_update();
1449 "hcall_vphn() experienced a hardware fault "
1450 "preventing VPHN. Disabling polling...\n");
1451 stop_topology_update();
1458 * Update the node maps and sysfs entries for each cpu whose home node
1461 int arch_update_cpu_topology(void)
1463 int cpu, nid, old_nid;
1464 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1465 struct sys_device *sysdev;
1467 for_each_cpu(cpu,&cpu_associativity_changes_mask) {
1468 vphn_get_associativity(cpu, associativity);
1469 nid = associativity_to_nid(associativity);
1471 if (nid < 0 || !node_online(nid))
1472 nid = first_online_node;
1474 old_nid = numa_cpu_lookup_table[cpu];
1476 /* Disable hotplug while we update the cpu
1480 unregister_cpu_under_node(cpu, old_nid);
1481 unmap_cpu_from_node(cpu);
1482 map_cpu_to_node(cpu, nid);
1483 register_cpu_under_node(cpu, nid);
1486 sysdev = get_cpu_sysdev(cpu);
1488 kobject_uevent(&sysdev->kobj, KOBJ_CHANGE);
1494 static void topology_work_fn(struct work_struct *work)
1496 rebuild_sched_domains();
1498 static DECLARE_WORK(topology_work, topology_work_fn);
1500 void topology_schedule_update(void)
1502 schedule_work(&topology_work);
1505 static void topology_timer_fn(unsigned long ignored)
1509 if (update_cpu_associativity_changes_mask() > 0)
1510 topology_schedule_update();
1511 set_topology_timer();
1513 static struct timer_list topology_timer =
1514 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1516 static void set_topology_timer(void)
1518 topology_timer.data = 0;
1519 topology_timer.expires = jiffies + 60 * HZ;
1520 add_timer(&topology_timer);
1524 * Start polling for VPHN associativity changes.
1526 int start_topology_update(void)
1530 /* Disabled until races with load balancing are fixed */
1531 if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1532 get_lppaca()->shared_proc) {
1534 setup_cpu_associativity_change_counters();
1535 init_timer_deferrable(&topology_timer);
1536 set_topology_timer();
1542 __initcall(start_topology_update);
1545 * Disable polling for VPHN associativity changes.
1547 int stop_topology_update(void)
1550 return del_timer_sync(&topology_timer);
1552 #endif /* CONFIG_PPC_SPLPAR */