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 <linux/stop_machine.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/uaccess.h>
29 #include <linux/slab.h>
30 #include <asm/cputhreads.h>
31 #include <asm/sparsemem.h>
34 #include <asm/firmware.h>
36 #include <asm/hvcall.h>
37 #include <asm/setup.h>
40 static int numa_enabled = 1;
42 static char *cmdline __initdata;
44 static int numa_debug;
45 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
47 int numa_cpu_lookup_table[NR_CPUS];
48 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
49 struct pglist_data *node_data[MAX_NUMNODES];
51 EXPORT_SYMBOL(numa_cpu_lookup_table);
52 EXPORT_SYMBOL(node_to_cpumask_map);
53 EXPORT_SYMBOL(node_data);
55 static int min_common_depth;
56 static int n_mem_addr_cells, n_mem_size_cells;
57 static int form1_affinity;
59 #define MAX_DISTANCE_REF_POINTS 4
60 static int distance_ref_points_depth;
61 static const unsigned int *distance_ref_points;
62 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
65 * Allocate node_to_cpumask_map based on number of available nodes
66 * Requires node_possible_map to be valid.
68 * Note: cpumask_of_node() is not valid until after this is done.
70 static void __init setup_node_to_cpumask_map(void)
74 /* setup nr_node_ids if not done yet */
75 if (nr_node_ids == MAX_NUMNODES)
78 /* allocate the map */
79 for (node = 0; node < nr_node_ids; node++)
80 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
82 /* cpumask_of_node() will now work */
83 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
86 static int __init fake_numa_create_new_node(unsigned long end_pfn,
89 unsigned long long mem;
91 static unsigned int fake_nid;
92 static unsigned long long curr_boundary;
95 * Modify node id, iff we started creating NUMA nodes
96 * We want to continue from where we left of the last time
101 * In case there are no more arguments to parse, the
102 * node_id should be the same as the last fake node id
103 * (we've handled this above).
108 mem = memparse(p, &p);
112 if (mem < curr_boundary)
117 if ((end_pfn << PAGE_SHIFT) > mem) {
119 * Skip commas and spaces
121 while (*p == ',' || *p == ' ' || *p == '\t')
127 dbg("created new fake_node with id %d\n", fake_nid);
134 * get_node_active_region - Return active region containing pfn
135 * Active range returned is empty if none found.
136 * @pfn: The page to return the region for
137 * @node_ar: Returned set to the active region containing @pfn
139 static void __init get_node_active_region(unsigned long pfn,
140 struct node_active_region *node_ar)
142 unsigned long start_pfn, end_pfn;
145 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
146 if (pfn >= start_pfn && pfn < end_pfn) {
148 node_ar->start_pfn = start_pfn;
149 node_ar->end_pfn = end_pfn;
155 static void map_cpu_to_node(int cpu, int node)
157 numa_cpu_lookup_table[cpu] = node;
159 dbg("adding cpu %d to node %d\n", cpu, node);
161 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
162 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
165 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
166 static void unmap_cpu_from_node(unsigned long cpu)
168 int node = numa_cpu_lookup_table[cpu];
170 dbg("removing cpu %lu from node %d\n", cpu, node);
172 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
173 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
175 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
179 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
181 /* must hold reference to node during call */
182 static const int *of_get_associativity(struct device_node *dev)
184 return of_get_property(dev, "ibm,associativity", NULL);
188 * Returns the property linux,drconf-usable-memory if
189 * it exists (the property exists only in kexec/kdump kernels,
190 * added by kexec-tools)
192 static const u32 *of_get_usable_memory(struct device_node *memory)
196 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
197 if (!prop || len < sizeof(unsigned int))
202 int __node_distance(int a, int b)
205 int distance = LOCAL_DISTANCE;
208 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
210 for (i = 0; i < distance_ref_points_depth; i++) {
211 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
214 /* Double the distance for each NUMA level */
221 static void initialize_distance_lookup_table(int nid,
222 const unsigned int *associativity)
229 for (i = 0; i < distance_ref_points_depth; i++) {
230 distance_lookup_table[nid][i] =
231 associativity[distance_ref_points[i]];
235 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
238 static int associativity_to_nid(const unsigned int *associativity)
242 if (min_common_depth == -1)
245 if (associativity[0] >= min_common_depth)
246 nid = associativity[min_common_depth];
248 /* POWER4 LPAR uses 0xffff as invalid node */
249 if (nid == 0xffff || nid >= MAX_NUMNODES)
252 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
253 initialize_distance_lookup_table(nid, associativity);
259 /* Returns the nid associated with the given device tree node,
260 * or -1 if not found.
262 static int of_node_to_nid_single(struct device_node *device)
265 const unsigned int *tmp;
267 tmp = of_get_associativity(device);
269 nid = associativity_to_nid(tmp);
273 /* Walk the device tree upwards, looking for an associativity id */
274 int of_node_to_nid(struct device_node *device)
276 struct device_node *tmp;
281 nid = of_node_to_nid_single(device);
286 device = of_get_parent(tmp);
293 EXPORT_SYMBOL_GPL(of_node_to_nid);
295 static int __init find_min_common_depth(void)
298 struct device_node *root;
300 if (firmware_has_feature(FW_FEATURE_OPAL))
301 root = of_find_node_by_path("/ibm,opal");
303 root = of_find_node_by_path("/rtas");
305 root = of_find_node_by_path("/");
308 * This property is a set of 32-bit integers, each representing
309 * an index into the ibm,associativity nodes.
311 * With form 0 affinity the first integer is for an SMP configuration
312 * (should be all 0's) and the second is for a normal NUMA
313 * configuration. We have only one level of NUMA.
315 * With form 1 affinity the first integer is the most significant
316 * NUMA boundary and the following are progressively less significant
317 * boundaries. There can be more than one level of NUMA.
319 distance_ref_points = of_get_property(root,
320 "ibm,associativity-reference-points",
321 &distance_ref_points_depth);
323 if (!distance_ref_points) {
324 dbg("NUMA: ibm,associativity-reference-points not found.\n");
328 distance_ref_points_depth /= sizeof(int);
330 if (firmware_has_feature(FW_FEATURE_OPAL) ||
331 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
332 dbg("Using form 1 affinity\n");
336 if (form1_affinity) {
337 depth = distance_ref_points[0];
339 if (distance_ref_points_depth < 2) {
340 printk(KERN_WARNING "NUMA: "
341 "short ibm,associativity-reference-points\n");
345 depth = distance_ref_points[1];
349 * Warn and cap if the hardware supports more than
350 * MAX_DISTANCE_REF_POINTS domains.
352 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
353 printk(KERN_WARNING "NUMA: distance array capped at "
354 "%d entries\n", MAX_DISTANCE_REF_POINTS);
355 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
366 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
368 struct device_node *memory = NULL;
370 memory = of_find_node_by_type(memory, "memory");
372 panic("numa.c: No memory nodes found!");
374 *n_addr_cells = of_n_addr_cells(memory);
375 *n_size_cells = of_n_size_cells(memory);
379 static unsigned long read_n_cells(int n, const unsigned int **buf)
381 unsigned long result = 0;
384 result = (result << 32) | **buf;
391 * Read the next memblock list entry from the ibm,dynamic-memory property
392 * and return the information in the provided of_drconf_cell structure.
394 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
398 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
401 drmem->drc_index = cp[0];
402 drmem->reserved = cp[1];
403 drmem->aa_index = cp[2];
404 drmem->flags = cp[3];
410 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
412 * The layout of the ibm,dynamic-memory property is a number N of memblock
413 * list entries followed by N memblock list entries. Each memblock list entry
414 * contains information as laid out in the of_drconf_cell struct above.
416 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
421 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
422 if (!prop || len < sizeof(unsigned int))
427 /* Now that we know the number of entries, revalidate the size
428 * of the property read in to ensure we have everything
430 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
438 * Retrieve and validate the ibm,lmb-size property for drconf memory
439 * from the device tree.
441 static u64 of_get_lmb_size(struct device_node *memory)
446 prop = of_get_property(memory, "ibm,lmb-size", &len);
447 if (!prop || len < sizeof(unsigned int))
450 return read_n_cells(n_mem_size_cells, &prop);
453 struct assoc_arrays {
460 * Retrieve and validate the list of associativity arrays for drconf
461 * memory from the ibm,associativity-lookup-arrays property of the
464 * The layout of the ibm,associativity-lookup-arrays property is a number N
465 * indicating the number of associativity arrays, followed by a number M
466 * indicating the size of each associativity array, followed by a list
467 * of N associativity arrays.
469 static int of_get_assoc_arrays(struct device_node *memory,
470 struct assoc_arrays *aa)
475 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
476 if (!prop || len < 2 * sizeof(unsigned int))
479 aa->n_arrays = *prop++;
480 aa->array_sz = *prop++;
482 /* Now that we know the number of arrays and size of each array,
483 * revalidate the size of the property read in.
485 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
493 * This is like of_node_to_nid_single() for memory represented in the
494 * ibm,dynamic-reconfiguration-memory node.
496 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
497 struct assoc_arrays *aa)
500 int nid = default_nid;
503 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
504 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
505 drmem->aa_index < aa->n_arrays) {
506 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
507 nid = aa->arrays[index];
509 if (nid == 0xffff || nid >= MAX_NUMNODES)
517 * Figure out to which domain a cpu belongs and stick it there.
518 * Return the id of the domain used.
520 static int numa_setup_cpu(unsigned long lcpu)
523 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
530 nid = of_node_to_nid_single(cpu);
532 if (nid < 0 || !node_online(nid))
533 nid = first_online_node;
535 map_cpu_to_node(lcpu, nid);
542 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
545 unsigned long lcpu = (unsigned long)hcpu;
546 int ret = NOTIFY_DONE;
550 case CPU_UP_PREPARE_FROZEN:
551 numa_setup_cpu(lcpu);
554 #ifdef CONFIG_HOTPLUG_CPU
556 case CPU_DEAD_FROZEN:
557 case CPU_UP_CANCELED:
558 case CPU_UP_CANCELED_FROZEN:
559 unmap_cpu_from_node(lcpu);
568 * Check and possibly modify a memory region to enforce the memory limit.
570 * Returns the size the region should have to enforce the memory limit.
571 * This will either be the original value of size, a truncated value,
572 * or zero. If the returned value of size is 0 the region should be
573 * discarded as it lies wholly above the memory limit.
575 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
579 * We use memblock_end_of_DRAM() in here instead of memory_limit because
580 * we've already adjusted it for the limit and it takes care of
581 * having memory holes below the limit. Also, in the case of
582 * iommu_is_off, memory_limit is not set but is implicitly enforced.
585 if (start + size <= memblock_end_of_DRAM())
588 if (start >= memblock_end_of_DRAM())
591 return memblock_end_of_DRAM() - start;
595 * Reads the counter for a given entry in
596 * linux,drconf-usable-memory property
598 static inline int __init read_usm_ranges(const u32 **usm)
601 * For each lmb in ibm,dynamic-memory a corresponding
602 * entry in linux,drconf-usable-memory property contains
603 * a counter followed by that many (base, size) duple.
604 * read the counter from linux,drconf-usable-memory
606 return read_n_cells(n_mem_size_cells, usm);
610 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
611 * node. This assumes n_mem_{addr,size}_cells have been set.
613 static void __init parse_drconf_memory(struct device_node *memory)
615 const u32 *uninitialized_var(dm), *usm;
616 unsigned int n, rc, ranges, is_kexec_kdump = 0;
617 unsigned long lmb_size, base, size, sz;
619 struct assoc_arrays aa = { .arrays = NULL };
621 n = of_get_drconf_memory(memory, &dm);
625 lmb_size = of_get_lmb_size(memory);
629 rc = of_get_assoc_arrays(memory, &aa);
633 /* check if this is a kexec/kdump kernel */
634 usm = of_get_usable_memory(memory);
638 for (; n != 0; --n) {
639 struct of_drconf_cell drmem;
641 read_drconf_cell(&drmem, &dm);
643 /* skip this block if the reserved bit is set in flags (0x80)
644 or if the block is not assigned to this partition (0x8) */
645 if ((drmem.flags & DRCONF_MEM_RESERVED)
646 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
649 base = drmem.base_addr;
653 if (is_kexec_kdump) {
654 ranges = read_usm_ranges(&usm);
655 if (!ranges) /* there are no (base, size) duple */
659 if (is_kexec_kdump) {
660 base = read_n_cells(n_mem_addr_cells, &usm);
661 size = read_n_cells(n_mem_size_cells, &usm);
663 nid = of_drconf_to_nid_single(&drmem, &aa);
664 fake_numa_create_new_node(
665 ((base + size) >> PAGE_SHIFT),
667 node_set_online(nid);
668 sz = numa_enforce_memory_limit(base, size);
670 memblock_set_node(base, sz, nid);
675 static int __init parse_numa_properties(void)
677 struct device_node *memory;
681 if (numa_enabled == 0) {
682 printk(KERN_WARNING "NUMA disabled by user\n");
686 min_common_depth = find_min_common_depth();
688 if (min_common_depth < 0)
689 return min_common_depth;
691 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
694 * Even though we connect cpus to numa domains later in SMP
695 * init, we need to know the node ids now. This is because
696 * each node to be onlined must have NODE_DATA etc backing it.
698 for_each_present_cpu(i) {
699 struct device_node *cpu;
702 cpu = of_get_cpu_node(i, NULL);
704 nid = of_node_to_nid_single(cpu);
708 * Don't fall back to default_nid yet -- we will plug
709 * cpus into nodes once the memory scan has discovered
714 node_set_online(nid);
717 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
719 for_each_node_by_type(memory, "memory") {
724 const unsigned int *memcell_buf;
727 memcell_buf = of_get_property(memory,
728 "linux,usable-memory", &len);
729 if (!memcell_buf || len <= 0)
730 memcell_buf = of_get_property(memory, "reg", &len);
731 if (!memcell_buf || len <= 0)
735 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
737 /* these are order-sensitive, and modify the buffer pointer */
738 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
739 size = read_n_cells(n_mem_size_cells, &memcell_buf);
742 * Assumption: either all memory nodes or none will
743 * have associativity properties. If none, then
744 * everything goes to default_nid.
746 nid = of_node_to_nid_single(memory);
750 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
751 node_set_online(nid);
753 if (!(size = numa_enforce_memory_limit(start, size))) {
760 memblock_set_node(start, size, nid);
767 * Now do the same thing for each MEMBLOCK listed in the
768 * ibm,dynamic-memory property in the
769 * ibm,dynamic-reconfiguration-memory node.
771 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
773 parse_drconf_memory(memory);
778 static void __init setup_nonnuma(void)
780 unsigned long top_of_ram = memblock_end_of_DRAM();
781 unsigned long total_ram = memblock_phys_mem_size();
782 unsigned long start_pfn, end_pfn;
783 unsigned int nid = 0;
784 struct memblock_region *reg;
786 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
787 top_of_ram, total_ram);
788 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
789 (top_of_ram - total_ram) >> 20);
791 for_each_memblock(memory, reg) {
792 start_pfn = memblock_region_memory_base_pfn(reg);
793 end_pfn = memblock_region_memory_end_pfn(reg);
795 fake_numa_create_new_node(end_pfn, &nid);
796 memblock_set_node(PFN_PHYS(start_pfn),
797 PFN_PHYS(end_pfn - start_pfn), nid);
798 node_set_online(nid);
802 void __init dump_numa_cpu_topology(void)
805 unsigned int cpu, count;
807 if (min_common_depth == -1 || !numa_enabled)
810 for_each_online_node(node) {
811 printk(KERN_DEBUG "Node %d CPUs:", node);
815 * If we used a CPU iterator here we would miss printing
816 * the holes in the cpumap.
818 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
819 if (cpumask_test_cpu(cpu,
820 node_to_cpumask_map[node])) {
826 printk("-%u", cpu - 1);
832 printk("-%u", nr_cpu_ids - 1);
837 static void __init dump_numa_memory_topology(void)
842 if (min_common_depth == -1 || !numa_enabled)
845 for_each_online_node(node) {
848 printk(KERN_DEBUG "Node %d Memory:", node);
852 for (i = 0; i < memblock_end_of_DRAM();
853 i += (1 << SECTION_SIZE_BITS)) {
854 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
872 * Allocate some memory, satisfying the memblock or bootmem allocator where
873 * required. nid is the preferred node and end is the physical address of
874 * the highest address in the node.
876 * Returns the virtual address of the memory.
878 static void __init *careful_zallocation(int nid, unsigned long size,
880 unsigned long end_pfn)
884 unsigned long ret_paddr;
886 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
888 /* retry over all memory */
890 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
893 panic("numa.c: cannot allocate %lu bytes for node %d",
896 ret = __va(ret_paddr);
899 * We initialize the nodes in numeric order: 0, 1, 2...
900 * and hand over control from the MEMBLOCK allocator to the
901 * bootmem allocator. If this function is called for
902 * node 5, then we know that all nodes <5 are using the
903 * bootmem allocator instead of the MEMBLOCK allocator.
905 * So, check the nid from which this allocation came
906 * and double check to see if we need to use bootmem
907 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
908 * since it would be useless.
910 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
912 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
915 dbg("alloc_bootmem %p %lx\n", ret, size);
918 memset(ret, 0, size);
922 static struct notifier_block ppc64_numa_nb = {
923 .notifier_call = cpu_numa_callback,
924 .priority = 1 /* Must run before sched domains notifier. */
927 static void __init mark_reserved_regions_for_nid(int nid)
929 struct pglist_data *node = NODE_DATA(nid);
930 struct memblock_region *reg;
932 for_each_memblock(reserved, reg) {
933 unsigned long physbase = reg->base;
934 unsigned long size = reg->size;
935 unsigned long start_pfn = physbase >> PAGE_SHIFT;
936 unsigned long end_pfn = PFN_UP(physbase + size);
937 struct node_active_region node_ar;
938 unsigned long node_end_pfn = node->node_start_pfn +
939 node->node_spanned_pages;
942 * Check to make sure that this memblock.reserved area is
943 * within the bounds of the node that we care about.
944 * Checking the nid of the start and end points is not
945 * sufficient because the reserved area could span the
948 if (end_pfn <= node->node_start_pfn ||
949 start_pfn >= node_end_pfn)
952 get_node_active_region(start_pfn, &node_ar);
953 while (start_pfn < end_pfn &&
954 node_ar.start_pfn < node_ar.end_pfn) {
955 unsigned long reserve_size = size;
957 * if reserved region extends past active region
958 * then trim size to active region
960 if (end_pfn > node_ar.end_pfn)
961 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
964 * Only worry about *this* node, others may not
965 * yet have valid NODE_DATA().
967 if (node_ar.nid == nid) {
968 dbg("reserve_bootmem %lx %lx nid=%d\n",
969 physbase, reserve_size, node_ar.nid);
970 reserve_bootmem_node(NODE_DATA(node_ar.nid),
971 physbase, reserve_size,
975 * if reserved region is contained in the active region
978 if (end_pfn <= node_ar.end_pfn)
982 * reserved region extends past the active region
983 * get next active region that contains this
986 start_pfn = node_ar.end_pfn;
987 physbase = start_pfn << PAGE_SHIFT;
988 size = size - reserve_size;
989 get_node_active_region(start_pfn, &node_ar);
995 void __init do_init_bootmem(void)
1000 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1001 max_pfn = max_low_pfn;
1003 if (parse_numa_properties())
1006 dump_numa_memory_topology();
1008 for_each_online_node(nid) {
1009 unsigned long start_pfn, end_pfn;
1010 void *bootmem_vaddr;
1011 unsigned long bootmap_pages;
1013 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1016 * Allocate the node structure node local if possible
1018 * Be careful moving this around, as it relies on all
1019 * previous nodes' bootmem to be initialized and have
1020 * all reserved areas marked.
1022 NODE_DATA(nid) = careful_zallocation(nid,
1023 sizeof(struct pglist_data),
1024 SMP_CACHE_BYTES, end_pfn);
1026 dbg("node %d\n", nid);
1027 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1029 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1030 NODE_DATA(nid)->node_start_pfn = start_pfn;
1031 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1033 if (NODE_DATA(nid)->node_spanned_pages == 0)
1036 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1037 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1039 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1040 bootmem_vaddr = careful_zallocation(nid,
1041 bootmap_pages << PAGE_SHIFT,
1042 PAGE_SIZE, end_pfn);
1044 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1046 init_bootmem_node(NODE_DATA(nid),
1047 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1048 start_pfn, end_pfn);
1050 free_bootmem_with_active_regions(nid, end_pfn);
1052 * Be very careful about moving this around. Future
1053 * calls to careful_zallocation() depend on this getting
1056 mark_reserved_regions_for_nid(nid);
1057 sparse_memory_present_with_active_regions(nid);
1060 init_bootmem_done = 1;
1063 * Now bootmem is initialised we can create the node to cpumask
1064 * lookup tables and setup the cpu callback to populate them.
1066 setup_node_to_cpumask_map();
1068 register_cpu_notifier(&ppc64_numa_nb);
1069 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1070 (void *)(unsigned long)boot_cpuid);
1073 void __init paging_init(void)
1075 unsigned long max_zone_pfns[MAX_NR_ZONES];
1076 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1077 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1078 free_area_init_nodes(max_zone_pfns);
1081 static int __init early_numa(char *p)
1086 if (strstr(p, "off"))
1089 if (strstr(p, "debug"))
1092 p = strstr(p, "fake=");
1094 cmdline = p + strlen("fake=");
1098 early_param("numa", early_numa);
1100 #ifdef CONFIG_MEMORY_HOTPLUG
1102 * Find the node associated with a hot added memory section for
1103 * memory represented in the device tree by the property
1104 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1106 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1107 unsigned long scn_addr)
1110 unsigned int drconf_cell_cnt, rc;
1111 unsigned long lmb_size;
1112 struct assoc_arrays aa;
1115 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1116 if (!drconf_cell_cnt)
1119 lmb_size = of_get_lmb_size(memory);
1123 rc = of_get_assoc_arrays(memory, &aa);
1127 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1128 struct of_drconf_cell drmem;
1130 read_drconf_cell(&drmem, &dm);
1132 /* skip this block if it is reserved or not assigned to
1134 if ((drmem.flags & DRCONF_MEM_RESERVED)
1135 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1138 if ((scn_addr < drmem.base_addr)
1139 || (scn_addr >= (drmem.base_addr + lmb_size)))
1142 nid = of_drconf_to_nid_single(&drmem, &aa);
1150 * Find the node associated with a hot added memory section for memory
1151 * represented in the device tree as a node (i.e. memory@XXXX) for
1154 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1156 struct device_node *memory;
1159 for_each_node_by_type(memory, "memory") {
1160 unsigned long start, size;
1162 const unsigned int *memcell_buf;
1165 memcell_buf = of_get_property(memory, "reg", &len);
1166 if (!memcell_buf || len <= 0)
1169 /* ranges in cell */
1170 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1173 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1174 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1176 if ((scn_addr < start) || (scn_addr >= (start + size)))
1179 nid = of_node_to_nid_single(memory);
1187 of_node_put(memory);
1193 * Find the node associated with a hot added memory section. Section
1194 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1195 * sections are fully contained within a single MEMBLOCK.
1197 int hot_add_scn_to_nid(unsigned long scn_addr)
1199 struct device_node *memory = NULL;
1202 if (!numa_enabled || (min_common_depth < 0))
1203 return first_online_node;
1205 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1207 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1208 of_node_put(memory);
1210 nid = hot_add_node_scn_to_nid(scn_addr);
1213 if (nid < 0 || !node_online(nid))
1214 nid = first_online_node;
1216 if (NODE_DATA(nid)->node_spanned_pages)
1219 for_each_online_node(nid) {
1220 if (NODE_DATA(nid)->node_spanned_pages) {
1230 static u64 hot_add_drconf_memory_max(void)
1232 struct device_node *memory = NULL;
1233 unsigned int drconf_cell_cnt = 0;
1237 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1239 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1240 lmb_size = of_get_lmb_size(memory);
1241 of_node_put(memory);
1243 return lmb_size * drconf_cell_cnt;
1247 * memory_hotplug_max - return max address of memory that may be added
1249 * This is currently only used on systems that support drconfig memory
1252 u64 memory_hotplug_max(void)
1254 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1256 #endif /* CONFIG_MEMORY_HOTPLUG */
1258 /* Virtual Processor Home Node (VPHN) support */
1259 #ifdef CONFIG_PPC_SPLPAR
1260 struct topology_update_data {
1261 struct topology_update_data *next;
1267 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1268 static cpumask_t cpu_associativity_changes_mask;
1269 static int vphn_enabled;
1270 static int prrn_enabled;
1271 static void reset_topology_timer(void);
1274 * Store the current values of the associativity change counters in the
1277 static void setup_cpu_associativity_change_counters(void)
1281 /* The VPHN feature supports a maximum of 8 reference points */
1282 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1284 for_each_possible_cpu(cpu) {
1286 u8 *counts = vphn_cpu_change_counts[cpu];
1287 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1289 for (i = 0; i < distance_ref_points_depth; i++)
1290 counts[i] = hypervisor_counts[i];
1295 * The hypervisor maintains a set of 8 associativity change counters in
1296 * the VPA of each cpu that correspond to the associativity levels in the
1297 * ibm,associativity-reference-points property. When an associativity
1298 * level changes, the corresponding counter is incremented.
1300 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1301 * node associativity levels have changed.
1303 * Returns the number of cpus with unhandled associativity changes.
1305 static int update_cpu_associativity_changes_mask(void)
1308 cpumask_t *changes = &cpu_associativity_changes_mask;
1310 for_each_possible_cpu(cpu) {
1312 u8 *counts = vphn_cpu_change_counts[cpu];
1313 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1315 for (i = 0; i < distance_ref_points_depth; i++) {
1316 if (hypervisor_counts[i] != counts[i]) {
1317 counts[i] = hypervisor_counts[i];
1322 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1323 cpu = cpu_last_thread_sibling(cpu);
1327 return cpumask_weight(changes);
1331 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1332 * the complete property we have to add the length in the first cell.
1334 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1337 * Convert the associativity domain numbers returned from the hypervisor
1338 * to the sequence they would appear in the ibm,associativity property.
1340 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1342 int i, nr_assoc_doms = 0;
1343 const u16 *field = (const u16*) packed;
1345 #define VPHN_FIELD_UNUSED (0xffff)
1346 #define VPHN_FIELD_MSB (0x8000)
1347 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1349 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1350 if (*field == VPHN_FIELD_UNUSED) {
1351 /* All significant fields processed, and remaining
1352 * fields contain the reserved value of all 1's.
1355 unpacked[i] = *((u32*)field);
1357 } else if (*field & VPHN_FIELD_MSB) {
1358 /* Data is in the lower 15 bits of this field */
1359 unpacked[i] = *field & VPHN_FIELD_MASK;
1363 /* Data is in the lower 15 bits of this field
1364 * concatenated with the next 16 bit field
1366 unpacked[i] = *((u32*)field);
1372 /* The first cell contains the length of the property */
1373 unpacked[0] = nr_assoc_doms;
1375 return nr_assoc_doms;
1379 * Retrieve the new associativity information for a virtual processor's
1382 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1385 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1387 int hwcpu = get_hard_smp_processor_id(cpu);
1389 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1390 vphn_unpack_associativity(retbuf, associativity);
1395 static long vphn_get_associativity(unsigned long cpu,
1396 unsigned int *associativity)
1400 rc = hcall_vphn(cpu, associativity);
1405 "VPHN is not supported. Disabling polling...\n");
1406 stop_topology_update();
1410 "hcall_vphn() experienced a hardware fault "
1411 "preventing VPHN. Disabling polling...\n");
1412 stop_topology_update();
1419 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1420 * characteristics change. This function doesn't perform any locking and is
1421 * only safe to call from stop_machine().
1423 static int update_cpu_topology(void *data)
1425 struct topology_update_data *update;
1431 cpu = smp_processor_id();
1433 for (update = data; update; update = update->next) {
1434 if (cpu != update->cpu)
1437 unmap_cpu_from_node(update->cpu);
1438 map_cpu_to_node(update->cpu, update->new_nid);
1446 * Update the node maps and sysfs entries for each cpu whose home node
1447 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1449 int arch_update_cpu_topology(void)
1451 unsigned int cpu, sibling, changed = 0;
1452 struct topology_update_data *updates, *ud;
1453 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1454 cpumask_t updated_cpus;
1456 int weight, new_nid, i = 0;
1458 weight = cpumask_weight(&cpu_associativity_changes_mask);
1462 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1466 cpumask_clear(&updated_cpus);
1468 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1470 * If siblings aren't flagged for changes, updates list
1471 * will be too short. Skip on this update and set for next
1474 if (!cpumask_subset(cpu_sibling_mask(cpu),
1475 &cpu_associativity_changes_mask)) {
1476 pr_info("Sibling bits not set for associativity "
1477 "change, cpu%d\n", cpu);
1478 cpumask_or(&cpu_associativity_changes_mask,
1479 &cpu_associativity_changes_mask,
1480 cpu_sibling_mask(cpu));
1481 cpu = cpu_last_thread_sibling(cpu);
1485 /* Use associativity from first thread for all siblings */
1486 vphn_get_associativity(cpu, associativity);
1487 new_nid = associativity_to_nid(associativity);
1488 if (new_nid < 0 || !node_online(new_nid))
1489 new_nid = first_online_node;
1491 if (new_nid == numa_cpu_lookup_table[cpu]) {
1492 cpumask_andnot(&cpu_associativity_changes_mask,
1493 &cpu_associativity_changes_mask,
1494 cpu_sibling_mask(cpu));
1495 cpu = cpu_last_thread_sibling(cpu);
1499 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1502 ud->new_nid = new_nid;
1503 ud->old_nid = numa_cpu_lookup_table[sibling];
1504 cpumask_set_cpu(sibling, &updated_cpus);
1506 ud->next = &updates[i];
1508 cpu = cpu_last_thread_sibling(cpu);
1511 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1513 for (ud = &updates[0]; ud; ud = ud->next) {
1514 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1515 register_cpu_under_node(ud->cpu, ud->new_nid);
1517 dev = get_cpu_device(ud->cpu);
1519 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1520 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1528 static void topology_work_fn(struct work_struct *work)
1530 rebuild_sched_domains();
1532 static DECLARE_WORK(topology_work, topology_work_fn);
1534 void topology_schedule_update(void)
1536 schedule_work(&topology_work);
1539 static void topology_timer_fn(unsigned long ignored)
1541 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1542 topology_schedule_update();
1543 else if (vphn_enabled) {
1544 if (update_cpu_associativity_changes_mask() > 0)
1545 topology_schedule_update();
1546 reset_topology_timer();
1549 static struct timer_list topology_timer =
1550 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1552 static void reset_topology_timer(void)
1554 topology_timer.data = 0;
1555 topology_timer.expires = jiffies + 60 * HZ;
1556 mod_timer(&topology_timer, topology_timer.expires);
1561 static void stage_topology_update(int core_id)
1563 cpumask_or(&cpu_associativity_changes_mask,
1564 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1565 reset_topology_timer();
1568 static int dt_update_callback(struct notifier_block *nb,
1569 unsigned long action, void *data)
1571 struct of_prop_reconfig *update;
1572 int rc = NOTIFY_DONE;
1575 case OF_RECONFIG_UPDATE_PROPERTY:
1576 update = (struct of_prop_reconfig *)data;
1577 if (!of_prop_cmp(update->dn->type, "cpu") &&
1578 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1580 of_property_read_u32(update->dn, "reg", &core_id);
1581 stage_topology_update(core_id);
1590 static struct notifier_block dt_update_nb = {
1591 .notifier_call = dt_update_callback,
1597 * Start polling for associativity changes.
1599 int start_topology_update(void)
1603 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1604 if (!prrn_enabled) {
1608 rc = of_reconfig_notifier_register(&dt_update_nb);
1611 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1612 get_lppaca()->shared_proc) {
1613 if (!vphn_enabled) {
1616 setup_cpu_associativity_change_counters();
1617 init_timer_deferrable(&topology_timer);
1618 reset_topology_timer();
1626 * Disable polling for VPHN associativity changes.
1628 int stop_topology_update(void)
1635 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1637 } else if (vphn_enabled) {
1639 rc = del_timer_sync(&topology_timer);
1645 int prrn_is_enabled(void)
1647 return prrn_enabled;
1650 static int topology_read(struct seq_file *file, void *v)
1652 if (vphn_enabled || prrn_enabled)
1653 seq_puts(file, "on\n");
1655 seq_puts(file, "off\n");
1660 static int topology_open(struct inode *inode, struct file *file)
1662 return single_open(file, topology_read, NULL);
1665 static ssize_t topology_write(struct file *file, const char __user *buf,
1666 size_t count, loff_t *off)
1668 char kbuf[4]; /* "on" or "off" plus null. */
1671 read_len = count < 3 ? count : 3;
1672 if (copy_from_user(kbuf, buf, read_len))
1675 kbuf[read_len] = '\0';
1677 if (!strncmp(kbuf, "on", 2))
1678 start_topology_update();
1679 else if (!strncmp(kbuf, "off", 3))
1680 stop_topology_update();
1687 static const struct file_operations topology_ops = {
1689 .write = topology_write,
1690 .open = topology_open,
1691 .release = single_release
1694 static int topology_update_init(void)
1696 start_topology_update();
1697 proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1701 device_initcall(topology_update_init);
1702 #endif /* CONFIG_PPC_SPLPAR */