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/slab.h>
26 #include <asm/sparsemem.h>
29 #include <asm/firmware.h>
31 #include <asm/hvcall.h>
32 #include <asm/setup.h>
34 static int numa_enabled = 1;
36 static char *cmdline __initdata;
38 static int numa_debug;
39 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
41 int numa_cpu_lookup_table[NR_CPUS];
42 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
43 struct pglist_data *node_data[MAX_NUMNODES];
45 EXPORT_SYMBOL(numa_cpu_lookup_table);
46 EXPORT_SYMBOL(node_to_cpumask_map);
47 EXPORT_SYMBOL(node_data);
49 static int min_common_depth;
50 static int n_mem_addr_cells, n_mem_size_cells;
51 static int form1_affinity;
53 #define MAX_DISTANCE_REF_POINTS 4
54 static int distance_ref_points_depth;
55 static const unsigned int *distance_ref_points;
56 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
59 * Allocate node_to_cpumask_map based on number of available nodes
60 * Requires node_possible_map to be valid.
62 * Note: cpumask_of_node() is not valid until after this is done.
64 static void __init setup_node_to_cpumask_map(void)
68 /* setup nr_node_ids if not done yet */
69 if (nr_node_ids == MAX_NUMNODES)
72 /* allocate the map */
73 for (node = 0; node < nr_node_ids; node++)
74 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
76 /* cpumask_of_node() will now work */
77 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
80 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
83 unsigned long long mem;
85 static unsigned int fake_nid;
86 static unsigned long long curr_boundary;
89 * Modify node id, iff we started creating NUMA nodes
90 * We want to continue from where we left of the last time
95 * In case there are no more arguments to parse, the
96 * node_id should be the same as the last fake node id
97 * (we've handled this above).
102 mem = memparse(p, &p);
106 if (mem < curr_boundary)
111 if ((end_pfn << PAGE_SHIFT) > mem) {
113 * Skip commas and spaces
115 while (*p == ',' || *p == ' ' || *p == '\t')
121 dbg("created new fake_node with id %d\n", fake_nid);
128 * get_node_active_region - Return active region containing pfn
129 * Active range returned is empty if none found.
130 * @pfn: The page to return the region for
131 * @node_ar: Returned set to the active region containing @pfn
133 static void __init get_node_active_region(unsigned long pfn,
134 struct node_active_region *node_ar)
136 unsigned long start_pfn, end_pfn;
139 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
140 if (pfn >= start_pfn && pfn < end_pfn) {
142 node_ar->start_pfn = start_pfn;
143 node_ar->end_pfn = end_pfn;
149 static void map_cpu_to_node(int cpu, int node)
151 numa_cpu_lookup_table[cpu] = node;
153 dbg("adding cpu %d to node %d\n", cpu, node);
155 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
156 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
159 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
160 static void unmap_cpu_from_node(unsigned long cpu)
162 int node = numa_cpu_lookup_table[cpu];
164 dbg("removing cpu %lu from node %d\n", cpu, node);
166 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
167 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
169 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
173 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
175 /* must hold reference to node during call */
176 static const int *of_get_associativity(struct device_node *dev)
178 return of_get_property(dev, "ibm,associativity", NULL);
182 * Returns the property linux,drconf-usable-memory if
183 * it exists (the property exists only in kexec/kdump kernels,
184 * added by kexec-tools)
186 static const u32 *of_get_usable_memory(struct device_node *memory)
190 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
191 if (!prop || len < sizeof(unsigned int))
196 int __node_distance(int a, int b)
199 int distance = LOCAL_DISTANCE;
204 for (i = 0; i < distance_ref_points_depth; i++) {
205 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
208 /* Double the distance for each NUMA level */
215 static void initialize_distance_lookup_table(int nid,
216 const unsigned int *associativity)
223 for (i = 0; i < distance_ref_points_depth; i++) {
224 distance_lookup_table[nid][i] =
225 associativity[distance_ref_points[i]];
229 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
232 static int associativity_to_nid(const unsigned int *associativity)
236 if (min_common_depth == -1)
239 if (associativity[0] >= min_common_depth)
240 nid = associativity[min_common_depth];
242 /* POWER4 LPAR uses 0xffff as invalid node */
243 if (nid == 0xffff || nid >= MAX_NUMNODES)
246 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
247 initialize_distance_lookup_table(nid, associativity);
253 /* Returns the nid associated with the given device tree node,
254 * or -1 if not found.
256 static int of_node_to_nid_single(struct device_node *device)
259 const unsigned int *tmp;
261 tmp = of_get_associativity(device);
263 nid = associativity_to_nid(tmp);
267 /* Walk the device tree upwards, looking for an associativity id */
268 int of_node_to_nid(struct device_node *device)
270 struct device_node *tmp;
275 nid = of_node_to_nid_single(device);
280 device = of_get_parent(tmp);
287 EXPORT_SYMBOL_GPL(of_node_to_nid);
289 static int __init find_min_common_depth(void)
292 struct device_node *chosen;
293 struct device_node *root;
296 if (firmware_has_feature(FW_FEATURE_OPAL))
297 root = of_find_node_by_path("/ibm,opal");
299 root = of_find_node_by_path("/rtas");
301 root = of_find_node_by_path("/");
304 * This property is a set of 32-bit integers, each representing
305 * an index into the ibm,associativity nodes.
307 * With form 0 affinity the first integer is for an SMP configuration
308 * (should be all 0's) and the second is for a normal NUMA
309 * configuration. We have only one level of NUMA.
311 * With form 1 affinity the first integer is the most significant
312 * NUMA boundary and the following are progressively less significant
313 * boundaries. There can be more than one level of NUMA.
315 distance_ref_points = of_get_property(root,
316 "ibm,associativity-reference-points",
317 &distance_ref_points_depth);
319 if (!distance_ref_points) {
320 dbg("NUMA: ibm,associativity-reference-points not found.\n");
324 distance_ref_points_depth /= sizeof(int);
326 #define VEC5_AFFINITY_BYTE 5
327 #define VEC5_AFFINITY 0x80
329 if (firmware_has_feature(FW_FEATURE_OPAL))
332 chosen = of_find_node_by_path("/chosen");
334 vec5 = of_get_property(chosen,
335 "ibm,architecture-vec-5", NULL);
336 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] &
338 dbg("Using form 1 affinity\n");
346 if (form1_affinity) {
347 depth = distance_ref_points[0];
349 if (distance_ref_points_depth < 2) {
350 printk(KERN_WARNING "NUMA: "
351 "short ibm,associativity-reference-points\n");
355 depth = distance_ref_points[1];
359 * Warn and cap if the hardware supports more than
360 * MAX_DISTANCE_REF_POINTS domains.
362 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
363 printk(KERN_WARNING "NUMA: distance array capped at "
364 "%d entries\n", MAX_DISTANCE_REF_POINTS);
365 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
376 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
378 struct device_node *memory = NULL;
380 memory = of_find_node_by_type(memory, "memory");
382 panic("numa.c: No memory nodes found!");
384 *n_addr_cells = of_n_addr_cells(memory);
385 *n_size_cells = of_n_size_cells(memory);
389 static unsigned long read_n_cells(int n, const unsigned int **buf)
391 unsigned long result = 0;
394 result = (result << 32) | **buf;
401 * Read the next memblock list entry from the ibm,dynamic-memory property
402 * and return the information in the provided of_drconf_cell structure.
404 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
408 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
411 drmem->drc_index = cp[0];
412 drmem->reserved = cp[1];
413 drmem->aa_index = cp[2];
414 drmem->flags = cp[3];
420 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
422 * The layout of the ibm,dynamic-memory property is a number N of memblock
423 * list entries followed by N memblock list entries. Each memblock list entry
424 * contains information as laid out in the of_drconf_cell struct above.
426 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
431 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
432 if (!prop || len < sizeof(unsigned int))
437 /* Now that we know the number of entries, revalidate the size
438 * of the property read in to ensure we have everything
440 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
448 * Retrieve and validate the ibm,lmb-size property for drconf memory
449 * from the device tree.
451 static u64 of_get_lmb_size(struct device_node *memory)
456 prop = of_get_property(memory, "ibm,lmb-size", &len);
457 if (!prop || len < sizeof(unsigned int))
460 return read_n_cells(n_mem_size_cells, &prop);
463 struct assoc_arrays {
470 * Retrieve and validate the list of associativity arrays for drconf
471 * memory from the ibm,associativity-lookup-arrays property of the
474 * The layout of the ibm,associativity-lookup-arrays property is a number N
475 * indicating the number of associativity arrays, followed by a number M
476 * indicating the size of each associativity array, followed by a list
477 * of N associativity arrays.
479 static int of_get_assoc_arrays(struct device_node *memory,
480 struct assoc_arrays *aa)
485 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
486 if (!prop || len < 2 * sizeof(unsigned int))
489 aa->n_arrays = *prop++;
490 aa->array_sz = *prop++;
492 /* Now that we know the number of arrays and size of each array,
493 * revalidate the size of the property read in.
495 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
503 * This is like of_node_to_nid_single() for memory represented in the
504 * ibm,dynamic-reconfiguration-memory node.
506 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
507 struct assoc_arrays *aa)
510 int nid = default_nid;
513 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
514 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
515 drmem->aa_index < aa->n_arrays) {
516 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
517 nid = aa->arrays[index];
519 if (nid == 0xffff || nid >= MAX_NUMNODES)
527 * Figure out to which domain a cpu belongs and stick it there.
528 * Return the id of the domain used.
530 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
533 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
540 nid = of_node_to_nid_single(cpu);
542 if (nid < 0 || !node_online(nid))
543 nid = first_online_node;
545 map_cpu_to_node(lcpu, nid);
552 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
553 unsigned long action,
556 unsigned long lcpu = (unsigned long)hcpu;
557 int ret = NOTIFY_DONE;
561 case CPU_UP_PREPARE_FROZEN:
562 numa_setup_cpu(lcpu);
565 #ifdef CONFIG_HOTPLUG_CPU
567 case CPU_DEAD_FROZEN:
568 case CPU_UP_CANCELED:
569 case CPU_UP_CANCELED_FROZEN:
570 unmap_cpu_from_node(lcpu);
579 * Check and possibly modify a memory region to enforce the memory limit.
581 * Returns the size the region should have to enforce the memory limit.
582 * This will either be the original value of size, a truncated value,
583 * or zero. If the returned value of size is 0 the region should be
584 * discarded as it lies wholly above the memory limit.
586 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
590 * We use memblock_end_of_DRAM() in here instead of memory_limit because
591 * we've already adjusted it for the limit and it takes care of
592 * having memory holes below the limit. Also, in the case of
593 * iommu_is_off, memory_limit is not set but is implicitly enforced.
596 if (start + size <= memblock_end_of_DRAM())
599 if (start >= memblock_end_of_DRAM())
602 return memblock_end_of_DRAM() - start;
606 * Reads the counter for a given entry in
607 * linux,drconf-usable-memory property
609 static inline int __init read_usm_ranges(const u32 **usm)
612 * For each lmb in ibm,dynamic-memory a corresponding
613 * entry in linux,drconf-usable-memory property contains
614 * a counter followed by that many (base, size) duple.
615 * read the counter from linux,drconf-usable-memory
617 return read_n_cells(n_mem_size_cells, usm);
621 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
622 * node. This assumes n_mem_{addr,size}_cells have been set.
624 static void __init parse_drconf_memory(struct device_node *memory)
626 const u32 *uninitialized_var(dm), *usm;
627 unsigned int n, rc, ranges, is_kexec_kdump = 0;
628 unsigned long lmb_size, base, size, sz;
630 struct assoc_arrays aa = { .arrays = NULL };
632 n = of_get_drconf_memory(memory, &dm);
636 lmb_size = of_get_lmb_size(memory);
640 rc = of_get_assoc_arrays(memory, &aa);
644 /* check if this is a kexec/kdump kernel */
645 usm = of_get_usable_memory(memory);
649 for (; n != 0; --n) {
650 struct of_drconf_cell drmem;
652 read_drconf_cell(&drmem, &dm);
654 /* skip this block if the reserved bit is set in flags (0x80)
655 or if the block is not assigned to this partition (0x8) */
656 if ((drmem.flags & DRCONF_MEM_RESERVED)
657 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
660 base = drmem.base_addr;
664 if (is_kexec_kdump) {
665 ranges = read_usm_ranges(&usm);
666 if (!ranges) /* there are no (base, size) duple */
670 if (is_kexec_kdump) {
671 base = read_n_cells(n_mem_addr_cells, &usm);
672 size = read_n_cells(n_mem_size_cells, &usm);
674 nid = of_drconf_to_nid_single(&drmem, &aa);
675 fake_numa_create_new_node(
676 ((base + size) >> PAGE_SHIFT),
678 node_set_online(nid);
679 sz = numa_enforce_memory_limit(base, size);
681 memblock_set_node(base, sz, nid);
686 static int __init parse_numa_properties(void)
688 struct device_node *memory;
692 if (numa_enabled == 0) {
693 printk(KERN_WARNING "NUMA disabled by user\n");
697 min_common_depth = find_min_common_depth();
699 if (min_common_depth < 0)
700 return min_common_depth;
702 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
705 * Even though we connect cpus to numa domains later in SMP
706 * init, we need to know the node ids now. This is because
707 * each node to be onlined must have NODE_DATA etc backing it.
709 for_each_present_cpu(i) {
710 struct device_node *cpu;
713 cpu = of_get_cpu_node(i, NULL);
715 nid = of_node_to_nid_single(cpu);
719 * Don't fall back to default_nid yet -- we will plug
720 * cpus into nodes once the memory scan has discovered
725 node_set_online(nid);
728 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
730 for_each_node_by_type(memory, "memory") {
735 const unsigned int *memcell_buf;
738 memcell_buf = of_get_property(memory,
739 "linux,usable-memory", &len);
740 if (!memcell_buf || len <= 0)
741 memcell_buf = of_get_property(memory, "reg", &len);
742 if (!memcell_buf || len <= 0)
746 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
748 /* these are order-sensitive, and modify the buffer pointer */
749 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
750 size = read_n_cells(n_mem_size_cells, &memcell_buf);
753 * Assumption: either all memory nodes or none will
754 * have associativity properties. If none, then
755 * everything goes to default_nid.
757 nid = of_node_to_nid_single(memory);
761 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
762 node_set_online(nid);
764 if (!(size = numa_enforce_memory_limit(start, size))) {
771 memblock_set_node(start, size, nid);
778 * Now do the same thing for each MEMBLOCK listed in the
779 * ibm,dynamic-memory property in the
780 * ibm,dynamic-reconfiguration-memory node.
782 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
784 parse_drconf_memory(memory);
789 static void __init setup_nonnuma(void)
791 unsigned long top_of_ram = memblock_end_of_DRAM();
792 unsigned long total_ram = memblock_phys_mem_size();
793 unsigned long start_pfn, end_pfn;
794 unsigned int nid = 0;
795 struct memblock_region *reg;
797 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
798 top_of_ram, total_ram);
799 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
800 (top_of_ram - total_ram) >> 20);
802 for_each_memblock(memory, reg) {
803 start_pfn = memblock_region_memory_base_pfn(reg);
804 end_pfn = memblock_region_memory_end_pfn(reg);
806 fake_numa_create_new_node(end_pfn, &nid);
807 memblock_set_node(PFN_PHYS(start_pfn),
808 PFN_PHYS(end_pfn - start_pfn), nid);
809 node_set_online(nid);
813 void __init dump_numa_cpu_topology(void)
816 unsigned int cpu, count;
818 if (min_common_depth == -1 || !numa_enabled)
821 for_each_online_node(node) {
822 printk(KERN_DEBUG "Node %d CPUs:", node);
826 * If we used a CPU iterator here we would miss printing
827 * the holes in the cpumap.
829 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
830 if (cpumask_test_cpu(cpu,
831 node_to_cpumask_map[node])) {
837 printk("-%u", cpu - 1);
843 printk("-%u", nr_cpu_ids - 1);
848 static void __init dump_numa_memory_topology(void)
853 if (min_common_depth == -1 || !numa_enabled)
856 for_each_online_node(node) {
859 printk(KERN_DEBUG "Node %d Memory:", node);
863 for (i = 0; i < memblock_end_of_DRAM();
864 i += (1 << SECTION_SIZE_BITS)) {
865 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
883 * Allocate some memory, satisfying the memblock or bootmem allocator where
884 * required. nid is the preferred node and end is the physical address of
885 * the highest address in the node.
887 * Returns the virtual address of the memory.
889 static void __init *careful_zallocation(int nid, unsigned long size,
891 unsigned long end_pfn)
895 unsigned long ret_paddr;
897 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
899 /* retry over all memory */
901 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
904 panic("numa.c: cannot allocate %lu bytes for node %d",
907 ret = __va(ret_paddr);
910 * We initialize the nodes in numeric order: 0, 1, 2...
911 * and hand over control from the MEMBLOCK allocator to the
912 * bootmem allocator. If this function is called for
913 * node 5, then we know that all nodes <5 are using the
914 * bootmem allocator instead of the MEMBLOCK allocator.
916 * So, check the nid from which this allocation came
917 * and double check to see if we need to use bootmem
918 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
919 * since it would be useless.
921 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
923 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
926 dbg("alloc_bootmem %p %lx\n", ret, size);
929 memset(ret, 0, size);
933 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
934 .notifier_call = cpu_numa_callback,
935 .priority = 1 /* Must run before sched domains notifier. */
938 static void __init mark_reserved_regions_for_nid(int nid)
940 struct pglist_data *node = NODE_DATA(nid);
941 struct memblock_region *reg;
943 for_each_memblock(reserved, reg) {
944 unsigned long physbase = reg->base;
945 unsigned long size = reg->size;
946 unsigned long start_pfn = physbase >> PAGE_SHIFT;
947 unsigned long end_pfn = PFN_UP(physbase + size);
948 struct node_active_region node_ar;
949 unsigned long node_end_pfn = node->node_start_pfn +
950 node->node_spanned_pages;
953 * Check to make sure that this memblock.reserved area is
954 * within the bounds of the node that we care about.
955 * Checking the nid of the start and end points is not
956 * sufficient because the reserved area could span the
959 if (end_pfn <= node->node_start_pfn ||
960 start_pfn >= node_end_pfn)
963 get_node_active_region(start_pfn, &node_ar);
964 while (start_pfn < end_pfn &&
965 node_ar.start_pfn < node_ar.end_pfn) {
966 unsigned long reserve_size = size;
968 * if reserved region extends past active region
969 * then trim size to active region
971 if (end_pfn > node_ar.end_pfn)
972 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
975 * Only worry about *this* node, others may not
976 * yet have valid NODE_DATA().
978 if (node_ar.nid == nid) {
979 dbg("reserve_bootmem %lx %lx nid=%d\n",
980 physbase, reserve_size, node_ar.nid);
981 reserve_bootmem_node(NODE_DATA(node_ar.nid),
982 physbase, reserve_size,
986 * if reserved region is contained in the active region
989 if (end_pfn <= node_ar.end_pfn)
993 * reserved region extends past the active region
994 * get next active region that contains this
997 start_pfn = node_ar.end_pfn;
998 physbase = start_pfn << PAGE_SHIFT;
999 size = size - reserve_size;
1000 get_node_active_region(start_pfn, &node_ar);
1006 void __init do_init_bootmem(void)
1011 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1012 max_pfn = max_low_pfn;
1014 if (parse_numa_properties())
1017 dump_numa_memory_topology();
1019 for_each_online_node(nid) {
1020 unsigned long start_pfn, end_pfn;
1021 void *bootmem_vaddr;
1022 unsigned long bootmap_pages;
1024 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1027 * Allocate the node structure node local if possible
1029 * Be careful moving this around, as it relies on all
1030 * previous nodes' bootmem to be initialized and have
1031 * all reserved areas marked.
1033 NODE_DATA(nid) = careful_zallocation(nid,
1034 sizeof(struct pglist_data),
1035 SMP_CACHE_BYTES, end_pfn);
1037 dbg("node %d\n", nid);
1038 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1040 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1041 NODE_DATA(nid)->node_start_pfn = start_pfn;
1042 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1044 if (NODE_DATA(nid)->node_spanned_pages == 0)
1047 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1048 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1050 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1051 bootmem_vaddr = careful_zallocation(nid,
1052 bootmap_pages << PAGE_SHIFT,
1053 PAGE_SIZE, end_pfn);
1055 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1057 init_bootmem_node(NODE_DATA(nid),
1058 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1059 start_pfn, end_pfn);
1061 free_bootmem_with_active_regions(nid, end_pfn);
1063 * Be very careful about moving this around. Future
1064 * calls to careful_zallocation() depend on this getting
1067 mark_reserved_regions_for_nid(nid);
1068 sparse_memory_present_with_active_regions(nid);
1071 init_bootmem_done = 1;
1074 * Now bootmem is initialised we can create the node to cpumask
1075 * lookup tables and setup the cpu callback to populate them.
1077 setup_node_to_cpumask_map();
1079 register_cpu_notifier(&ppc64_numa_nb);
1080 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1081 (void *)(unsigned long)boot_cpuid);
1084 void __init paging_init(void)
1086 unsigned long max_zone_pfns[MAX_NR_ZONES];
1087 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1088 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1089 free_area_init_nodes(max_zone_pfns);
1092 static int __init early_numa(char *p)
1097 if (strstr(p, "off"))
1100 if (strstr(p, "debug"))
1103 p = strstr(p, "fake=");
1105 cmdline = p + strlen("fake=");
1109 early_param("numa", early_numa);
1111 #ifdef CONFIG_MEMORY_HOTPLUG
1113 * Find the node associated with a hot added memory section for
1114 * memory represented in the device tree by the property
1115 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1117 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1118 unsigned long scn_addr)
1121 unsigned int drconf_cell_cnt, rc;
1122 unsigned long lmb_size;
1123 struct assoc_arrays aa;
1126 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1127 if (!drconf_cell_cnt)
1130 lmb_size = of_get_lmb_size(memory);
1134 rc = of_get_assoc_arrays(memory, &aa);
1138 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1139 struct of_drconf_cell drmem;
1141 read_drconf_cell(&drmem, &dm);
1143 /* skip this block if it is reserved or not assigned to
1145 if ((drmem.flags & DRCONF_MEM_RESERVED)
1146 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1149 if ((scn_addr < drmem.base_addr)
1150 || (scn_addr >= (drmem.base_addr + lmb_size)))
1153 nid = of_drconf_to_nid_single(&drmem, &aa);
1161 * Find the node associated with a hot added memory section for memory
1162 * represented in the device tree as a node (i.e. memory@XXXX) for
1165 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1167 struct device_node *memory;
1170 for_each_node_by_type(memory, "memory") {
1171 unsigned long start, size;
1173 const unsigned int *memcell_buf;
1176 memcell_buf = of_get_property(memory, "reg", &len);
1177 if (!memcell_buf || len <= 0)
1180 /* ranges in cell */
1181 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1184 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1185 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1187 if ((scn_addr < start) || (scn_addr >= (start + size)))
1190 nid = of_node_to_nid_single(memory);
1198 of_node_put(memory);
1204 * Find the node associated with a hot added memory section. Section
1205 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1206 * sections are fully contained within a single MEMBLOCK.
1208 int hot_add_scn_to_nid(unsigned long scn_addr)
1210 struct device_node *memory = NULL;
1213 if (!numa_enabled || (min_common_depth < 0))
1214 return first_online_node;
1216 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1218 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1219 of_node_put(memory);
1221 nid = hot_add_node_scn_to_nid(scn_addr);
1224 if (nid < 0 || !node_online(nid))
1225 nid = first_online_node;
1227 if (NODE_DATA(nid)->node_spanned_pages)
1230 for_each_online_node(nid) {
1231 if (NODE_DATA(nid)->node_spanned_pages) {
1241 static u64 hot_add_drconf_memory_max(void)
1243 struct device_node *memory = NULL;
1244 unsigned int drconf_cell_cnt = 0;
1248 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1250 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1251 lmb_size = of_get_lmb_size(memory);
1252 of_node_put(memory);
1254 return lmb_size * drconf_cell_cnt;
1258 * memory_hotplug_max - return max address of memory that may be added
1260 * This is currently only used on systems that support drconfig memory
1263 u64 memory_hotplug_max(void)
1265 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1267 #endif /* CONFIG_MEMORY_HOTPLUG */
1269 /* Virtual Processor Home Node (VPHN) support */
1270 #ifdef CONFIG_PPC_SPLPAR
1271 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1272 static cpumask_t cpu_associativity_changes_mask;
1273 static int vphn_enabled;
1274 static void set_topology_timer(void);
1277 * Store the current values of the associativity change counters in the
1280 static void setup_cpu_associativity_change_counters(void)
1284 /* The VPHN feature supports a maximum of 8 reference points */
1285 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1287 for_each_possible_cpu(cpu) {
1289 u8 *counts = vphn_cpu_change_counts[cpu];
1290 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1292 for (i = 0; i < distance_ref_points_depth; i++)
1293 counts[i] = hypervisor_counts[i];
1298 * The hypervisor maintains a set of 8 associativity change counters in
1299 * the VPA of each cpu that correspond to the associativity levels in the
1300 * ibm,associativity-reference-points property. When an associativity
1301 * level changes, the corresponding counter is incremented.
1303 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1304 * node associativity levels have changed.
1306 * Returns the number of cpus with unhandled associativity changes.
1308 static int update_cpu_associativity_changes_mask(void)
1310 int cpu, nr_cpus = 0;
1311 cpumask_t *changes = &cpu_associativity_changes_mask;
1313 cpumask_clear(changes);
1315 for_each_possible_cpu(cpu) {
1317 u8 *counts = vphn_cpu_change_counts[cpu];
1318 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1320 for (i = 0; i < distance_ref_points_depth; i++) {
1321 if (hypervisor_counts[i] != counts[i]) {
1322 counts[i] = hypervisor_counts[i];
1327 cpumask_set_cpu(cpu, changes);
1336 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1337 * the complete property we have to add the length in the first cell.
1339 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1342 * Convert the associativity domain numbers returned from the hypervisor
1343 * to the sequence they would appear in the ibm,associativity property.
1345 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1347 int i, nr_assoc_doms = 0;
1348 const u16 *field = (const u16*) packed;
1350 #define VPHN_FIELD_UNUSED (0xffff)
1351 #define VPHN_FIELD_MSB (0x8000)
1352 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1354 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1355 if (*field == VPHN_FIELD_UNUSED) {
1356 /* All significant fields processed, and remaining
1357 * fields contain the reserved value of all 1's.
1360 unpacked[i] = *((u32*)field);
1362 } else if (*field & VPHN_FIELD_MSB) {
1363 /* Data is in the lower 15 bits of this field */
1364 unpacked[i] = *field & VPHN_FIELD_MASK;
1368 /* Data is in the lower 15 bits of this field
1369 * concatenated with the next 16 bit field
1371 unpacked[i] = *((u32*)field);
1377 /* The first cell contains the length of the property */
1378 unpacked[0] = nr_assoc_doms;
1380 return nr_assoc_doms;
1384 * Retrieve the new associativity information for a virtual processor's
1387 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1390 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1392 int hwcpu = get_hard_smp_processor_id(cpu);
1394 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1395 vphn_unpack_associativity(retbuf, associativity);
1400 static long vphn_get_associativity(unsigned long cpu,
1401 unsigned int *associativity)
1405 rc = hcall_vphn(cpu, associativity);
1410 "VPHN is not supported. Disabling polling...\n");
1411 stop_topology_update();
1415 "hcall_vphn() experienced a hardware fault "
1416 "preventing VPHN. Disabling polling...\n");
1417 stop_topology_update();
1424 * Update the node maps and sysfs entries for each cpu whose home node
1425 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1427 int arch_update_cpu_topology(void)
1429 int cpu, nid, old_nid, changed = 0;
1430 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1433 for_each_cpu(cpu,&cpu_associativity_changes_mask) {
1434 vphn_get_associativity(cpu, associativity);
1435 nid = associativity_to_nid(associativity);
1437 if (nid < 0 || !node_online(nid))
1438 nid = first_online_node;
1440 old_nid = numa_cpu_lookup_table[cpu];
1442 /* Disable hotplug while we update the cpu
1446 unregister_cpu_under_node(cpu, old_nid);
1447 unmap_cpu_from_node(cpu);
1448 map_cpu_to_node(cpu, nid);
1449 register_cpu_under_node(cpu, nid);
1452 dev = get_cpu_device(cpu);
1454 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1461 static void topology_work_fn(struct work_struct *work)
1463 rebuild_sched_domains();
1465 static DECLARE_WORK(topology_work, topology_work_fn);
1467 void topology_schedule_update(void)
1469 schedule_work(&topology_work);
1472 static void topology_timer_fn(unsigned long ignored)
1476 if (update_cpu_associativity_changes_mask() > 0)
1477 topology_schedule_update();
1478 set_topology_timer();
1480 static struct timer_list topology_timer =
1481 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1483 static void set_topology_timer(void)
1485 topology_timer.data = 0;
1486 topology_timer.expires = jiffies + 60 * HZ;
1487 add_timer(&topology_timer);
1491 * Start polling for VPHN associativity changes.
1493 int start_topology_update(void)
1497 /* Disabled until races with load balancing are fixed */
1498 if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1499 get_lppaca()->shared_proc) {
1501 setup_cpu_associativity_change_counters();
1502 init_timer_deferrable(&topology_timer);
1503 set_topology_timer();
1509 __initcall(start_topology_update);
1512 * Disable polling for VPHN associativity changes.
1514 int stop_topology_update(void)
1517 return del_timer_sync(&topology_timer);
1519 #endif /* CONFIG_PPC_SPLPAR */