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/module.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 <asm/sparsemem.h>
25 #include <asm/system.h>
28 static int numa_enabled = 1;
30 static char *cmdline __initdata;
32 static int numa_debug;
33 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
35 int numa_cpu_lookup_table[NR_CPUS];
36 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
37 struct pglist_data *node_data[MAX_NUMNODES];
39 EXPORT_SYMBOL(numa_cpu_lookup_table);
40 EXPORT_SYMBOL(node_to_cpumask_map);
41 EXPORT_SYMBOL(node_data);
43 static int min_common_depth;
44 static int n_mem_addr_cells, n_mem_size_cells;
45 static int form1_affinity;
47 #define MAX_DISTANCE_REF_POINTS 4
48 static int distance_ref_points_depth;
49 static const unsigned int *distance_ref_points;
50 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
53 * Allocate node_to_cpumask_map based on number of available nodes
54 * Requires node_possible_map to be valid.
56 * Note: node_to_cpumask() is not valid until after this is done.
58 static void __init setup_node_to_cpumask_map(void)
60 unsigned int node, num = 0;
62 /* setup nr_node_ids if not done yet */
63 if (nr_node_ids == MAX_NUMNODES) {
64 for_each_node_mask(node, node_possible_map)
66 nr_node_ids = num + 1;
69 /* allocate the map */
70 for (node = 0; node < nr_node_ids; node++)
71 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
73 /* cpumask_of_node() will now work */
74 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
77 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
80 unsigned long long mem;
82 static unsigned int fake_nid;
83 static unsigned long long curr_boundary;
86 * Modify node id, iff we started creating NUMA nodes
87 * We want to continue from where we left of the last time
92 * In case there are no more arguments to parse, the
93 * node_id should be the same as the last fake node id
94 * (we've handled this above).
99 mem = memparse(p, &p);
103 if (mem < curr_boundary)
108 if ((end_pfn << PAGE_SHIFT) > mem) {
110 * Skip commas and spaces
112 while (*p == ',' || *p == ' ' || *p == '\t')
118 dbg("created new fake_node with id %d\n", fake_nid);
125 * get_active_region_work_fn - A helper function for get_node_active_region
126 * Returns datax set to the start_pfn and end_pfn if they contain
127 * the initial value of datax->start_pfn between them
128 * @start_pfn: start page(inclusive) of region to check
129 * @end_pfn: end page(exclusive) of region to check
130 * @datax: comes in with ->start_pfn set to value to search for and
131 * goes out with active range if it contains it
132 * Returns 1 if search value is in range else 0
134 static int __init get_active_region_work_fn(unsigned long start_pfn,
135 unsigned long end_pfn, void *datax)
137 struct node_active_region *data;
138 data = (struct node_active_region *)datax;
140 if (start_pfn <= data->start_pfn && end_pfn > data->start_pfn) {
141 data->start_pfn = start_pfn;
142 data->end_pfn = end_pfn;
150 * get_node_active_region - Return active region containing start_pfn
151 * Active range returned is empty if none found.
152 * @start_pfn: The page to return the region for.
153 * @node_ar: Returned set to the active region containing start_pfn
155 static void __init get_node_active_region(unsigned long start_pfn,
156 struct node_active_region *node_ar)
158 int nid = early_pfn_to_nid(start_pfn);
161 node_ar->start_pfn = start_pfn;
162 node_ar->end_pfn = start_pfn;
163 work_with_active_regions(nid, get_active_region_work_fn, node_ar);
166 static void __cpuinit map_cpu_to_node(int cpu, int node)
168 numa_cpu_lookup_table[cpu] = node;
170 dbg("adding cpu %d to node %d\n", cpu, node);
172 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
173 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
176 #ifdef CONFIG_HOTPLUG_CPU
177 static void unmap_cpu_from_node(unsigned long cpu)
179 int node = numa_cpu_lookup_table[cpu];
181 dbg("removing cpu %lu from node %d\n", cpu, node);
183 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
184 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
186 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
190 #endif /* CONFIG_HOTPLUG_CPU */
192 /* must hold reference to node during call */
193 static const int *of_get_associativity(struct device_node *dev)
195 return of_get_property(dev, "ibm,associativity", NULL);
199 * Returns the property linux,drconf-usable-memory if
200 * it exists (the property exists only in kexec/kdump kernels,
201 * added by kexec-tools)
203 static const u32 *of_get_usable_memory(struct device_node *memory)
207 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
208 if (!prop || len < sizeof(unsigned int))
213 int __node_distance(int a, int b)
216 int distance = LOCAL_DISTANCE;
221 for (i = 0; i < distance_ref_points_depth; i++) {
222 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
225 /* Double the distance for each NUMA level */
232 static void initialize_distance_lookup_table(int nid,
233 const unsigned int *associativity)
240 for (i = 0; i < distance_ref_points_depth; i++) {
241 distance_lookup_table[nid][i] =
242 associativity[distance_ref_points[i]];
246 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
249 static int of_node_to_nid_single(struct device_node *device)
252 const unsigned int *tmp;
254 if (min_common_depth == -1)
257 tmp = of_get_associativity(device);
261 if (tmp[0] >= min_common_depth)
262 nid = tmp[min_common_depth];
264 /* POWER4 LPAR uses 0xffff as invalid node */
265 if (nid == 0xffff || nid >= MAX_NUMNODES)
268 if (nid > 0 && tmp[0] >= distance_ref_points_depth)
269 initialize_distance_lookup_table(nid, tmp);
275 /* Walk the device tree upwards, looking for an associativity id */
276 int of_node_to_nid(struct device_node *device)
278 struct device_node *tmp;
283 nid = of_node_to_nid_single(device);
288 device = of_get_parent(tmp);
295 EXPORT_SYMBOL_GPL(of_node_to_nid);
297 static int __init find_min_common_depth(void)
300 struct device_node *rtas_root;
301 struct device_node *chosen;
304 rtas_root = of_find_node_by_path("/rtas");
310 * This property is a set of 32-bit integers, each representing
311 * an index into the ibm,associativity nodes.
313 * With form 0 affinity the first integer is for an SMP configuration
314 * (should be all 0's) and the second is for a normal NUMA
315 * configuration. We have only one level of NUMA.
317 * With form 1 affinity the first integer is the most significant
318 * NUMA boundary and the following are progressively less significant
319 * boundaries. There can be more than one level of NUMA.
321 distance_ref_points = of_get_property(rtas_root,
322 "ibm,associativity-reference-points",
323 &distance_ref_points_depth);
325 if (!distance_ref_points) {
326 dbg("NUMA: ibm,associativity-reference-points not found.\n");
330 distance_ref_points_depth /= sizeof(int);
332 #define VEC5_AFFINITY_BYTE 5
333 #define VEC5_AFFINITY 0x80
334 chosen = of_find_node_by_path("/chosen");
336 vec5 = of_get_property(chosen, "ibm,architecture-vec-5", NULL);
337 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] & VEC5_AFFINITY)) {
338 dbg("Using form 1 affinity\n");
343 if (form1_affinity) {
344 depth = distance_ref_points[0];
346 if (distance_ref_points_depth < 2) {
347 printk(KERN_WARNING "NUMA: "
348 "short ibm,associativity-reference-points\n");
352 depth = distance_ref_points[1];
356 * Warn and cap if the hardware supports more than
357 * MAX_DISTANCE_REF_POINTS domains.
359 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
360 printk(KERN_WARNING "NUMA: distance array capped at "
361 "%d entries\n", MAX_DISTANCE_REF_POINTS);
362 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
365 of_node_put(rtas_root);
369 of_node_put(rtas_root);
373 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
375 struct device_node *memory = NULL;
377 memory = of_find_node_by_type(memory, "memory");
379 panic("numa.c: No memory nodes found!");
381 *n_addr_cells = of_n_addr_cells(memory);
382 *n_size_cells = of_n_size_cells(memory);
386 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
388 unsigned long result = 0;
391 result = (result << 32) | **buf;
397 struct of_drconf_cell {
405 #define DRCONF_MEM_ASSIGNED 0x00000008
406 #define DRCONF_MEM_AI_INVALID 0x00000040
407 #define DRCONF_MEM_RESERVED 0x00000080
410 * Read the next memblock list entry from the ibm,dynamic-memory property
411 * and return the information in the provided of_drconf_cell structure.
413 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
417 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
420 drmem->drc_index = cp[0];
421 drmem->reserved = cp[1];
422 drmem->aa_index = cp[2];
423 drmem->flags = cp[3];
429 * Retreive and validate the ibm,dynamic-memory property of the device tree.
431 * The layout of the ibm,dynamic-memory property is a number N of memblock
432 * list entries followed by N memblock list entries. Each memblock list entry
433 * contains information as layed out in the of_drconf_cell struct above.
435 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
440 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
441 if (!prop || len < sizeof(unsigned int))
446 /* Now that we know the number of entries, revalidate the size
447 * of the property read in to ensure we have everything
449 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
457 * Retreive and validate the ibm,lmb-size property for drconf memory
458 * from the device tree.
460 static u64 of_get_lmb_size(struct device_node *memory)
465 prop = of_get_property(memory, "ibm,lmb-size", &len);
466 if (!prop || len < sizeof(unsigned int))
469 return read_n_cells(n_mem_size_cells, &prop);
472 struct assoc_arrays {
479 * Retreive and validate the list of associativity arrays for drconf
480 * memory from the ibm,associativity-lookup-arrays property of the
483 * The layout of the ibm,associativity-lookup-arrays property is a number N
484 * indicating the number of associativity arrays, followed by a number M
485 * indicating the size of each associativity array, followed by a list
486 * of N associativity arrays.
488 static int of_get_assoc_arrays(struct device_node *memory,
489 struct assoc_arrays *aa)
494 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
495 if (!prop || len < 2 * sizeof(unsigned int))
498 aa->n_arrays = *prop++;
499 aa->array_sz = *prop++;
501 /* Now that we know the number of arrrays and size of each array,
502 * revalidate the size of the property read in.
504 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
512 * This is like of_node_to_nid_single() for memory represented in the
513 * ibm,dynamic-reconfiguration-memory node.
515 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
516 struct assoc_arrays *aa)
519 int nid = default_nid;
522 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
523 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
524 drmem->aa_index < aa->n_arrays) {
525 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
526 nid = aa->arrays[index];
528 if (nid == 0xffff || nid >= MAX_NUMNODES)
536 * Figure out to which domain a cpu belongs and stick it there.
537 * Return the id of the domain used.
539 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
542 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
549 nid = of_node_to_nid_single(cpu);
551 if (nid < 0 || !node_online(nid))
552 nid = first_online_node;
554 map_cpu_to_node(lcpu, nid);
561 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
562 unsigned long action,
565 unsigned long lcpu = (unsigned long)hcpu;
566 int ret = NOTIFY_DONE;
570 case CPU_UP_PREPARE_FROZEN:
571 numa_setup_cpu(lcpu);
574 #ifdef CONFIG_HOTPLUG_CPU
576 case CPU_DEAD_FROZEN:
577 case CPU_UP_CANCELED:
578 case CPU_UP_CANCELED_FROZEN:
579 unmap_cpu_from_node(lcpu);
588 * Check and possibly modify a memory region to enforce the memory limit.
590 * Returns the size the region should have to enforce the memory limit.
591 * This will either be the original value of size, a truncated value,
592 * or zero. If the returned value of size is 0 the region should be
593 * discarded as it lies wholy above the memory limit.
595 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
599 * We use memblock_end_of_DRAM() in here instead of memory_limit because
600 * we've already adjusted it for the limit and it takes care of
601 * having memory holes below the limit. Also, in the case of
602 * iommu_is_off, memory_limit is not set but is implicitly enforced.
605 if (start + size <= memblock_end_of_DRAM())
608 if (start >= memblock_end_of_DRAM())
611 return memblock_end_of_DRAM() - start;
615 * Reads the counter for a given entry in
616 * linux,drconf-usable-memory property
618 static inline int __init read_usm_ranges(const u32 **usm)
621 * For each lmb in ibm,dynamic-memory a corresponding
622 * entry in linux,drconf-usable-memory property contains
623 * a counter followed by that many (base, size) duple.
624 * read the counter from linux,drconf-usable-memory
626 return read_n_cells(n_mem_size_cells, usm);
630 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
631 * node. This assumes n_mem_{addr,size}_cells have been set.
633 static void __init parse_drconf_memory(struct device_node *memory)
636 unsigned int n, rc, ranges, is_kexec_kdump = 0;
637 unsigned long lmb_size, base, size, sz;
639 struct assoc_arrays aa;
641 n = of_get_drconf_memory(memory, &dm);
645 lmb_size = of_get_lmb_size(memory);
649 rc = of_get_assoc_arrays(memory, &aa);
653 /* check if this is a kexec/kdump kernel */
654 usm = of_get_usable_memory(memory);
658 for (; n != 0; --n) {
659 struct of_drconf_cell drmem;
661 read_drconf_cell(&drmem, &dm);
663 /* skip this block if the reserved bit is set in flags (0x80)
664 or if the block is not assigned to this partition (0x8) */
665 if ((drmem.flags & DRCONF_MEM_RESERVED)
666 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
669 base = drmem.base_addr;
673 if (is_kexec_kdump) {
674 ranges = read_usm_ranges(&usm);
675 if (!ranges) /* there are no (base, size) duple */
679 if (is_kexec_kdump) {
680 base = read_n_cells(n_mem_addr_cells, &usm);
681 size = read_n_cells(n_mem_size_cells, &usm);
683 nid = of_drconf_to_nid_single(&drmem, &aa);
684 fake_numa_create_new_node(
685 ((base + size) >> PAGE_SHIFT),
687 node_set_online(nid);
688 sz = numa_enforce_memory_limit(base, size);
690 add_active_range(nid, base >> PAGE_SHIFT,
692 + (sz >> PAGE_SHIFT));
697 static int __init parse_numa_properties(void)
699 struct device_node *cpu = NULL;
700 struct device_node *memory = NULL;
704 if (numa_enabled == 0) {
705 printk(KERN_WARNING "NUMA disabled by user\n");
709 min_common_depth = find_min_common_depth();
711 if (min_common_depth < 0)
712 return min_common_depth;
714 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
717 * Even though we connect cpus to numa domains later in SMP
718 * init, we need to know the node ids now. This is because
719 * each node to be onlined must have NODE_DATA etc backing it.
721 for_each_present_cpu(i) {
724 cpu = of_get_cpu_node(i, NULL);
726 nid = of_node_to_nid_single(cpu);
730 * Don't fall back to default_nid yet -- we will plug
731 * cpus into nodes once the memory scan has discovered
736 node_set_online(nid);
739 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
741 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
746 const unsigned int *memcell_buf;
749 memcell_buf = of_get_property(memory,
750 "linux,usable-memory", &len);
751 if (!memcell_buf || len <= 0)
752 memcell_buf = of_get_property(memory, "reg", &len);
753 if (!memcell_buf || len <= 0)
757 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
759 /* these are order-sensitive, and modify the buffer pointer */
760 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
761 size = read_n_cells(n_mem_size_cells, &memcell_buf);
764 * Assumption: either all memory nodes or none will
765 * have associativity properties. If none, then
766 * everything goes to default_nid.
768 nid = of_node_to_nid_single(memory);
772 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
773 node_set_online(nid);
775 if (!(size = numa_enforce_memory_limit(start, size))) {
782 add_active_range(nid, start >> PAGE_SHIFT,
783 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
790 * Now do the same thing for each MEMBLOCK listed in the ibm,dynamic-memory
791 * property in the ibm,dynamic-reconfiguration-memory node.
793 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
795 parse_drconf_memory(memory);
800 static void __init setup_nonnuma(void)
802 unsigned long top_of_ram = memblock_end_of_DRAM();
803 unsigned long total_ram = memblock_phys_mem_size();
804 unsigned long start_pfn, end_pfn;
805 unsigned int i, nid = 0;
807 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
808 top_of_ram, total_ram);
809 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
810 (top_of_ram - total_ram) >> 20);
812 for (i = 0; i < memblock.memory.cnt; ++i) {
813 start_pfn = memblock.memory.region[i].base >> PAGE_SHIFT;
814 end_pfn = start_pfn + memblock_size_pages(&memblock.memory, i);
816 fake_numa_create_new_node(end_pfn, &nid);
817 add_active_range(nid, start_pfn, end_pfn);
818 node_set_online(nid);
822 void __init dump_numa_cpu_topology(void)
825 unsigned int cpu, count;
827 if (min_common_depth == -1 || !numa_enabled)
830 for_each_online_node(node) {
831 printk(KERN_DEBUG "Node %d CPUs:", node);
835 * If we used a CPU iterator here we would miss printing
836 * the holes in the cpumap.
838 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
839 if (cpumask_test_cpu(cpu,
840 node_to_cpumask_map[node])) {
846 printk("-%u", cpu - 1);
852 printk("-%u", nr_cpu_ids - 1);
857 static void __init dump_numa_memory_topology(void)
862 if (min_common_depth == -1 || !numa_enabled)
865 for_each_online_node(node) {
868 printk(KERN_DEBUG "Node %d Memory:", node);
872 for (i = 0; i < memblock_end_of_DRAM();
873 i += (1 << SECTION_SIZE_BITS)) {
874 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
892 * Allocate some memory, satisfying the memblock or bootmem allocator where
893 * required. nid is the preferred node and end is the physical address of
894 * the highest address in the node.
896 * Returns the virtual address of the memory.
898 static void __init *careful_zallocation(int nid, unsigned long size,
900 unsigned long end_pfn)
904 unsigned long ret_paddr;
906 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
908 /* retry over all memory */
910 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
913 panic("numa.c: cannot allocate %lu bytes for node %d",
916 ret = __va(ret_paddr);
919 * We initialize the nodes in numeric order: 0, 1, 2...
920 * and hand over control from the MEMBLOCK allocator to the
921 * bootmem allocator. If this function is called for
922 * node 5, then we know that all nodes <5 are using the
923 * bootmem allocator instead of the MEMBLOCK allocator.
925 * So, check the nid from which this allocation came
926 * and double check to see if we need to use bootmem
927 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
928 * since it would be useless.
930 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
932 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
935 dbg("alloc_bootmem %p %lx\n", ret, size);
938 memset(ret, 0, size);
942 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
943 .notifier_call = cpu_numa_callback,
944 .priority = 1 /* Must run before sched domains notifier. */
947 static void mark_reserved_regions_for_nid(int nid)
949 struct pglist_data *node = NODE_DATA(nid);
952 for (i = 0; i < memblock.reserved.cnt; i++) {
953 unsigned long physbase = memblock.reserved.region[i].base;
954 unsigned long size = memblock.reserved.region[i].size;
955 unsigned long start_pfn = physbase >> PAGE_SHIFT;
956 unsigned long end_pfn = PFN_UP(physbase + size);
957 struct node_active_region node_ar;
958 unsigned long node_end_pfn = node->node_start_pfn +
959 node->node_spanned_pages;
962 * Check to make sure that this memblock.reserved area is
963 * within the bounds of the node that we care about.
964 * Checking the nid of the start and end points is not
965 * sufficient because the reserved area could span the
968 if (end_pfn <= node->node_start_pfn ||
969 start_pfn >= node_end_pfn)
972 get_node_active_region(start_pfn, &node_ar);
973 while (start_pfn < end_pfn &&
974 node_ar.start_pfn < node_ar.end_pfn) {
975 unsigned long reserve_size = size;
977 * if reserved region extends past active region
978 * then trim size to active region
980 if (end_pfn > node_ar.end_pfn)
981 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
984 * Only worry about *this* node, others may not
985 * yet have valid NODE_DATA().
987 if (node_ar.nid == nid) {
988 dbg("reserve_bootmem %lx %lx nid=%d\n",
989 physbase, reserve_size, node_ar.nid);
990 reserve_bootmem_node(NODE_DATA(node_ar.nid),
991 physbase, reserve_size,
995 * if reserved region is contained in the active region
998 if (end_pfn <= node_ar.end_pfn)
1002 * reserved region extends past the active region
1003 * get next active region that contains this
1006 start_pfn = node_ar.end_pfn;
1007 physbase = start_pfn << PAGE_SHIFT;
1008 size = size - reserve_size;
1009 get_node_active_region(start_pfn, &node_ar);
1015 void __init do_init_bootmem(void)
1020 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1021 max_pfn = max_low_pfn;
1023 if (parse_numa_properties())
1026 dump_numa_memory_topology();
1028 for_each_online_node(nid) {
1029 unsigned long start_pfn, end_pfn;
1030 void *bootmem_vaddr;
1031 unsigned long bootmap_pages;
1033 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1036 * Allocate the node structure node local if possible
1038 * Be careful moving this around, as it relies on all
1039 * previous nodes' bootmem to be initialized and have
1040 * all reserved areas marked.
1042 NODE_DATA(nid) = careful_zallocation(nid,
1043 sizeof(struct pglist_data),
1044 SMP_CACHE_BYTES, end_pfn);
1046 dbg("node %d\n", nid);
1047 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1049 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1050 NODE_DATA(nid)->node_start_pfn = start_pfn;
1051 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1053 if (NODE_DATA(nid)->node_spanned_pages == 0)
1056 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1057 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1059 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1060 bootmem_vaddr = careful_zallocation(nid,
1061 bootmap_pages << PAGE_SHIFT,
1062 PAGE_SIZE, end_pfn);
1064 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1066 init_bootmem_node(NODE_DATA(nid),
1067 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1068 start_pfn, end_pfn);
1070 free_bootmem_with_active_regions(nid, end_pfn);
1072 * Be very careful about moving this around. Future
1073 * calls to careful_zallocation() depend on this getting
1076 mark_reserved_regions_for_nid(nid);
1077 sparse_memory_present_with_active_regions(nid);
1080 init_bootmem_done = 1;
1083 * Now bootmem is initialised we can create the node to cpumask
1084 * lookup tables and setup the cpu callback to populate them.
1086 setup_node_to_cpumask_map();
1088 register_cpu_notifier(&ppc64_numa_nb);
1089 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1090 (void *)(unsigned long)boot_cpuid);
1093 void __init paging_init(void)
1095 unsigned long max_zone_pfns[MAX_NR_ZONES];
1096 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1097 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1098 free_area_init_nodes(max_zone_pfns);
1101 static int __init early_numa(char *p)
1106 if (strstr(p, "off"))
1109 if (strstr(p, "debug"))
1112 p = strstr(p, "fake=");
1114 cmdline = p + strlen("fake=");
1118 early_param("numa", early_numa);
1120 #ifdef CONFIG_MEMORY_HOTPLUG
1122 * Find the node associated with a hot added memory section for
1123 * memory represented in the device tree by the property
1124 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1126 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1127 unsigned long scn_addr)
1130 unsigned int drconf_cell_cnt, rc;
1131 unsigned long lmb_size;
1132 struct assoc_arrays aa;
1135 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1136 if (!drconf_cell_cnt)
1139 lmb_size = of_get_lmb_size(memory);
1143 rc = of_get_assoc_arrays(memory, &aa);
1147 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1148 struct of_drconf_cell drmem;
1150 read_drconf_cell(&drmem, &dm);
1152 /* skip this block if it is reserved or not assigned to
1154 if ((drmem.flags & DRCONF_MEM_RESERVED)
1155 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1158 if ((scn_addr < drmem.base_addr)
1159 || (scn_addr >= (drmem.base_addr + lmb_size)))
1162 nid = of_drconf_to_nid_single(&drmem, &aa);
1170 * Find the node associated with a hot added memory section for memory
1171 * represented in the device tree as a node (i.e. memory@XXXX) for
1174 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1176 struct device_node *memory = NULL;
1179 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
1180 unsigned long start, size;
1182 const unsigned int *memcell_buf;
1185 memcell_buf = of_get_property(memory, "reg", &len);
1186 if (!memcell_buf || len <= 0)
1189 /* ranges in cell */
1190 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1193 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1194 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1196 if ((scn_addr < start) || (scn_addr >= (start + size)))
1199 nid = of_node_to_nid_single(memory);
1203 of_node_put(memory);
1212 * Find the node associated with a hot added memory section. Section
1213 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1214 * sections are fully contained within a single MEMBLOCK.
1216 int hot_add_scn_to_nid(unsigned long scn_addr)
1218 struct device_node *memory = NULL;
1221 if (!numa_enabled || (min_common_depth < 0))
1222 return first_online_node;
1224 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1226 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1227 of_node_put(memory);
1229 nid = hot_add_node_scn_to_nid(scn_addr);
1232 if (nid < 0 || !node_online(nid))
1233 nid = first_online_node;
1235 if (NODE_DATA(nid)->node_spanned_pages)
1238 for_each_online_node(nid) {
1239 if (NODE_DATA(nid)->node_spanned_pages) {
1249 #endif /* CONFIG_MEMORY_HOTPLUG */