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/cputhreads.h>
35 #include <asm/topology.h>
36 #include <asm/firmware.h>
38 #include <asm/hvcall.h>
39 #include <asm/setup.h>
42 static int numa_enabled = 1;
44 static char *cmdline __initdata;
46 static int numa_debug;
47 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
49 int numa_cpu_lookup_table[NR_CPUS];
50 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
51 struct pglist_data *node_data[MAX_NUMNODES];
53 EXPORT_SYMBOL(numa_cpu_lookup_table);
54 EXPORT_SYMBOL(node_to_cpumask_map);
55 EXPORT_SYMBOL(node_data);
57 static int min_common_depth;
58 static int n_mem_addr_cells, n_mem_size_cells;
59 static int form1_affinity;
61 #define MAX_DISTANCE_REF_POINTS 4
62 static int distance_ref_points_depth;
63 static const __be32 *distance_ref_points;
64 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
67 * Allocate node_to_cpumask_map based on number of available nodes
68 * Requires node_possible_map to be valid.
70 * Note: cpumask_of_node() is not valid until after this is done.
72 static void __init setup_node_to_cpumask_map(void)
76 /* setup nr_node_ids if not done yet */
77 if (nr_node_ids == MAX_NUMNODES)
80 /* allocate the map */
81 for (node = 0; node < nr_node_ids; node++)
82 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
84 /* cpumask_of_node() will now work */
85 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
88 static int __init fake_numa_create_new_node(unsigned long end_pfn,
91 unsigned long long mem;
93 static unsigned int fake_nid;
94 static unsigned long long curr_boundary;
97 * Modify node id, iff we started creating NUMA nodes
98 * We want to continue from where we left of the last time
103 * In case there are no more arguments to parse, the
104 * node_id should be the same as the last fake node id
105 * (we've handled this above).
110 mem = memparse(p, &p);
114 if (mem < curr_boundary)
119 if ((end_pfn << PAGE_SHIFT) > mem) {
121 * Skip commas and spaces
123 while (*p == ',' || *p == ' ' || *p == '\t')
129 dbg("created new fake_node with id %d\n", fake_nid);
136 * get_node_active_region - Return active region containing pfn
137 * Active range returned is empty if none found.
138 * @pfn: The page to return the region for
139 * @node_ar: Returned set to the active region containing @pfn
141 static void __init get_node_active_region(unsigned long pfn,
142 struct node_active_region *node_ar)
144 unsigned long start_pfn, end_pfn;
147 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
148 if (pfn >= start_pfn && pfn < end_pfn) {
150 node_ar->start_pfn = start_pfn;
151 node_ar->end_pfn = end_pfn;
157 static void reset_numa_cpu_lookup_table(void)
161 for_each_possible_cpu(cpu)
162 numa_cpu_lookup_table[cpu] = -1;
165 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
167 numa_cpu_lookup_table[cpu] = node;
170 static void map_cpu_to_node(int cpu, int node)
172 update_numa_cpu_lookup_table(cpu, node);
174 dbg("adding cpu %d to node %d\n", cpu, node);
176 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
177 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
180 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
181 static void unmap_cpu_from_node(unsigned long cpu)
183 int node = numa_cpu_lookup_table[cpu];
185 dbg("removing cpu %lu from node %d\n", cpu, node);
187 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
188 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
190 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
194 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
196 /* must hold reference to node during call */
197 static const __be32 *of_get_associativity(struct device_node *dev)
199 return of_get_property(dev, "ibm,associativity", NULL);
203 * Returns the property linux,drconf-usable-memory if
204 * it exists (the property exists only in kexec/kdump kernels,
205 * added by kexec-tools)
207 static const __be32 *of_get_usable_memory(struct device_node *memory)
211 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
212 if (!prop || len < sizeof(unsigned int))
217 int __node_distance(int a, int b)
220 int distance = LOCAL_DISTANCE;
223 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
225 for (i = 0; i < distance_ref_points_depth; i++) {
226 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
229 /* Double the distance for each NUMA level */
236 static void initialize_distance_lookup_table(int nid,
237 const __be32 *associativity)
244 for (i = 0; i < distance_ref_points_depth; i++) {
247 entry = &associativity[be32_to_cpu(distance_ref_points[i])];
248 distance_lookup_table[nid][i] = of_read_number(entry, 1);
252 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
255 static int associativity_to_nid(const __be32 *associativity)
259 if (min_common_depth == -1)
262 if (of_read_number(associativity, 1) >= min_common_depth)
263 nid = of_read_number(&associativity[min_common_depth], 1);
265 /* POWER4 LPAR uses 0xffff as invalid node */
266 if (nid == 0xffff || nid >= MAX_NUMNODES)
270 of_read_number(associativity, 1) >= distance_ref_points_depth)
271 initialize_distance_lookup_table(nid, associativity);
277 /* Returns the nid associated with the given device tree node,
278 * or -1 if not found.
280 static int of_node_to_nid_single(struct device_node *device)
285 tmp = of_get_associativity(device);
287 nid = associativity_to_nid(tmp);
291 /* Walk the device tree upwards, looking for an associativity id */
292 int of_node_to_nid(struct device_node *device)
294 struct device_node *tmp;
299 nid = of_node_to_nid_single(device);
304 device = of_get_parent(tmp);
311 EXPORT_SYMBOL_GPL(of_node_to_nid);
313 static int __init find_min_common_depth(void)
316 struct device_node *root;
318 if (firmware_has_feature(FW_FEATURE_OPAL))
319 root = of_find_node_by_path("/ibm,opal");
321 root = of_find_node_by_path("/rtas");
323 root = of_find_node_by_path("/");
326 * This property is a set of 32-bit integers, each representing
327 * an index into the ibm,associativity nodes.
329 * With form 0 affinity the first integer is for an SMP configuration
330 * (should be all 0's) and the second is for a normal NUMA
331 * configuration. We have only one level of NUMA.
333 * With form 1 affinity the first integer is the most significant
334 * NUMA boundary and the following are progressively less significant
335 * boundaries. There can be more than one level of NUMA.
337 distance_ref_points = of_get_property(root,
338 "ibm,associativity-reference-points",
339 &distance_ref_points_depth);
341 if (!distance_ref_points) {
342 dbg("NUMA: ibm,associativity-reference-points not found.\n");
346 distance_ref_points_depth /= sizeof(int);
348 if (firmware_has_feature(FW_FEATURE_OPAL) ||
349 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
350 dbg("Using form 1 affinity\n");
354 if (form1_affinity) {
355 depth = of_read_number(distance_ref_points, 1);
357 if (distance_ref_points_depth < 2) {
358 printk(KERN_WARNING "NUMA: "
359 "short ibm,associativity-reference-points\n");
363 depth = of_read_number(&distance_ref_points[1], 1);
367 * Warn and cap if the hardware supports more than
368 * MAX_DISTANCE_REF_POINTS domains.
370 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
371 printk(KERN_WARNING "NUMA: distance array capped at "
372 "%d entries\n", MAX_DISTANCE_REF_POINTS);
373 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
384 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
386 struct device_node *memory = NULL;
388 memory = of_find_node_by_type(memory, "memory");
390 panic("numa.c: No memory nodes found!");
392 *n_addr_cells = of_n_addr_cells(memory);
393 *n_size_cells = of_n_size_cells(memory);
397 static unsigned long read_n_cells(int n, const __be32 **buf)
399 unsigned long result = 0;
402 result = (result << 32) | of_read_number(*buf, 1);
409 * Read the next memblock list entry from the ibm,dynamic-memory property
410 * and return the information in the provided of_drconf_cell structure.
412 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
416 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
419 drmem->drc_index = of_read_number(cp, 1);
420 drmem->reserved = of_read_number(&cp[1], 1);
421 drmem->aa_index = of_read_number(&cp[2], 1);
422 drmem->flags = of_read_number(&cp[3], 1);
428 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
430 * The layout of the ibm,dynamic-memory property is a number N of memblock
431 * list entries followed by N memblock list entries. Each memblock list entry
432 * contains information as laid out in the of_drconf_cell struct above.
434 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
439 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
440 if (!prop || len < sizeof(unsigned int))
443 entries = of_read_number(prop++, 1);
445 /* Now that we know the number of entries, revalidate the size
446 * of the property read in to ensure we have everything
448 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
456 * Retrieve and validate the ibm,lmb-size property for drconf memory
457 * from the device tree.
459 static u64 of_get_lmb_size(struct device_node *memory)
464 prop = of_get_property(memory, "ibm,lmb-size", &len);
465 if (!prop || len < sizeof(unsigned int))
468 return read_n_cells(n_mem_size_cells, &prop);
471 struct assoc_arrays {
474 const __be32 *arrays;
478 * Retrieve and validate the list of associativity arrays for drconf
479 * memory from the ibm,associativity-lookup-arrays property of the
482 * The layout of the ibm,associativity-lookup-arrays property is a number N
483 * indicating the number of associativity arrays, followed by a number M
484 * indicating the size of each associativity array, followed by a list
485 * of N associativity arrays.
487 static int of_get_assoc_arrays(struct device_node *memory,
488 struct assoc_arrays *aa)
493 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
494 if (!prop || len < 2 * sizeof(unsigned int))
497 aa->n_arrays = of_read_number(prop++, 1);
498 aa->array_sz = of_read_number(prop++, 1);
500 /* Now that we know the number of arrays and size of each array,
501 * revalidate the size of the property read in.
503 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
511 * This is like of_node_to_nid_single() for memory represented in the
512 * ibm,dynamic-reconfiguration-memory node.
514 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
515 struct assoc_arrays *aa)
518 int nid = default_nid;
521 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
522 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
523 drmem->aa_index < aa->n_arrays) {
524 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
525 nid = of_read_number(&aa->arrays[index], 1);
527 if (nid == 0xffff || nid >= MAX_NUMNODES)
535 * Figure out to which domain a cpu belongs and stick it there.
536 * Return the id of the domain used.
538 static int numa_setup_cpu(unsigned long lcpu)
541 struct device_node *cpu;
544 * If a valid cpu-to-node mapping is already available, use it
545 * directly instead of querying the firmware, since it represents
546 * the most recent mapping notified to us by the platform (eg: VPHN).
548 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
549 map_cpu_to_node(lcpu, nid);
553 cpu = of_get_cpu_node(lcpu, NULL);
561 nid = of_node_to_nid_single(cpu);
563 if (nid < 0 || !node_online(nid))
564 nid = first_online_node;
566 map_cpu_to_node(lcpu, nid);
573 static void verify_cpu_node_mapping(int cpu, int node)
575 int base, sibling, i;
577 /* Verify that all the threads in the core belong to the same node */
578 base = cpu_first_thread_sibling(cpu);
580 for (i = 0; i < threads_per_core; i++) {
583 if (sibling == cpu || cpu_is_offline(sibling))
586 if (cpu_to_node(sibling) != node) {
587 WARN(1, "CPU thread siblings %d and %d don't belong"
588 " to the same node!\n", cpu, sibling);
594 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
597 unsigned long lcpu = (unsigned long)hcpu;
598 int ret = NOTIFY_DONE, nid;
602 case CPU_UP_PREPARE_FROZEN:
603 nid = numa_setup_cpu(lcpu);
604 verify_cpu_node_mapping((int)lcpu, nid);
607 #ifdef CONFIG_HOTPLUG_CPU
609 case CPU_DEAD_FROZEN:
610 case CPU_UP_CANCELED:
611 case CPU_UP_CANCELED_FROZEN:
612 unmap_cpu_from_node(lcpu);
621 * Check and possibly modify a memory region to enforce the memory limit.
623 * Returns the size the region should have to enforce the memory limit.
624 * This will either be the original value of size, a truncated value,
625 * or zero. If the returned value of size is 0 the region should be
626 * discarded as it lies wholly above the memory limit.
628 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
632 * We use memblock_end_of_DRAM() in here instead of memory_limit because
633 * we've already adjusted it for the limit and it takes care of
634 * having memory holes below the limit. Also, in the case of
635 * iommu_is_off, memory_limit is not set but is implicitly enforced.
638 if (start + size <= memblock_end_of_DRAM())
641 if (start >= memblock_end_of_DRAM())
644 return memblock_end_of_DRAM() - start;
648 * Reads the counter for a given entry in
649 * linux,drconf-usable-memory property
651 static inline int __init read_usm_ranges(const __be32 **usm)
654 * For each lmb in ibm,dynamic-memory a corresponding
655 * entry in linux,drconf-usable-memory property contains
656 * a counter followed by that many (base, size) duple.
657 * read the counter from linux,drconf-usable-memory
659 return read_n_cells(n_mem_size_cells, usm);
663 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
664 * node. This assumes n_mem_{addr,size}_cells have been set.
666 static void __init parse_drconf_memory(struct device_node *memory)
668 const __be32 *uninitialized_var(dm), *usm;
669 unsigned int n, rc, ranges, is_kexec_kdump = 0;
670 unsigned long lmb_size, base, size, sz;
672 struct assoc_arrays aa = { .arrays = NULL };
674 n = of_get_drconf_memory(memory, &dm);
678 lmb_size = of_get_lmb_size(memory);
682 rc = of_get_assoc_arrays(memory, &aa);
686 /* check if this is a kexec/kdump kernel */
687 usm = of_get_usable_memory(memory);
691 for (; n != 0; --n) {
692 struct of_drconf_cell drmem;
694 read_drconf_cell(&drmem, &dm);
696 /* skip this block if the reserved bit is set in flags (0x80)
697 or if the block is not assigned to this partition (0x8) */
698 if ((drmem.flags & DRCONF_MEM_RESERVED)
699 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
702 base = drmem.base_addr;
706 if (is_kexec_kdump) {
707 ranges = read_usm_ranges(&usm);
708 if (!ranges) /* there are no (base, size) duple */
712 if (is_kexec_kdump) {
713 base = read_n_cells(n_mem_addr_cells, &usm);
714 size = read_n_cells(n_mem_size_cells, &usm);
716 nid = of_drconf_to_nid_single(&drmem, &aa);
717 fake_numa_create_new_node(
718 ((base + size) >> PAGE_SHIFT),
720 node_set_online(nid);
721 sz = numa_enforce_memory_limit(base, size);
723 memblock_set_node(base, sz,
724 &memblock.memory, nid);
729 static int __init parse_numa_properties(void)
731 struct device_node *memory;
735 if (numa_enabled == 0) {
736 printk(KERN_WARNING "NUMA disabled by user\n");
740 min_common_depth = find_min_common_depth();
742 if (min_common_depth < 0)
743 return min_common_depth;
745 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
748 * Even though we connect cpus to numa domains later in SMP
749 * init, we need to know the node ids now. This is because
750 * each node to be onlined must have NODE_DATA etc backing it.
752 for_each_present_cpu(i) {
753 struct device_node *cpu;
756 cpu = of_get_cpu_node(i, NULL);
758 nid = of_node_to_nid_single(cpu);
762 * Don't fall back to default_nid yet -- we will plug
763 * cpus into nodes once the memory scan has discovered
768 node_set_online(nid);
771 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
773 for_each_node_by_type(memory, "memory") {
778 const __be32 *memcell_buf;
781 memcell_buf = of_get_property(memory,
782 "linux,usable-memory", &len);
783 if (!memcell_buf || len <= 0)
784 memcell_buf = of_get_property(memory, "reg", &len);
785 if (!memcell_buf || len <= 0)
789 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
791 /* these are order-sensitive, and modify the buffer pointer */
792 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
793 size = read_n_cells(n_mem_size_cells, &memcell_buf);
796 * Assumption: either all memory nodes or none will
797 * have associativity properties. If none, then
798 * everything goes to default_nid.
800 nid = of_node_to_nid_single(memory);
804 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
805 node_set_online(nid);
807 if (!(size = numa_enforce_memory_limit(start, size))) {
814 memblock_set_node(start, size, &memblock.memory, nid);
821 * Now do the same thing for each MEMBLOCK listed in the
822 * ibm,dynamic-memory property in the
823 * ibm,dynamic-reconfiguration-memory node.
825 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
827 parse_drconf_memory(memory);
832 static void __init setup_nonnuma(void)
834 unsigned long top_of_ram = memblock_end_of_DRAM();
835 unsigned long total_ram = memblock_phys_mem_size();
836 unsigned long start_pfn, end_pfn;
837 unsigned int nid = 0;
838 struct memblock_region *reg;
840 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
841 top_of_ram, total_ram);
842 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
843 (top_of_ram - total_ram) >> 20);
845 for_each_memblock(memory, reg) {
846 start_pfn = memblock_region_memory_base_pfn(reg);
847 end_pfn = memblock_region_memory_end_pfn(reg);
849 fake_numa_create_new_node(end_pfn, &nid);
850 memblock_set_node(PFN_PHYS(start_pfn),
851 PFN_PHYS(end_pfn - start_pfn),
852 &memblock.memory, nid);
853 node_set_online(nid);
857 void __init dump_numa_cpu_topology(void)
860 unsigned int cpu, count;
862 if (min_common_depth == -1 || !numa_enabled)
865 for_each_online_node(node) {
866 printk(KERN_DEBUG "Node %d CPUs:", node);
870 * If we used a CPU iterator here we would miss printing
871 * the holes in the cpumap.
873 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
874 if (cpumask_test_cpu(cpu,
875 node_to_cpumask_map[node])) {
881 printk("-%u", cpu - 1);
887 printk("-%u", nr_cpu_ids - 1);
892 static void __init dump_numa_memory_topology(void)
897 if (min_common_depth == -1 || !numa_enabled)
900 for_each_online_node(node) {
903 printk(KERN_DEBUG "Node %d Memory:", node);
907 for (i = 0; i < memblock_end_of_DRAM();
908 i += (1 << SECTION_SIZE_BITS)) {
909 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
927 * Allocate some memory, satisfying the memblock or bootmem allocator where
928 * required. nid is the preferred node and end is the physical address of
929 * the highest address in the node.
931 * Returns the virtual address of the memory.
933 static void __init *careful_zallocation(int nid, unsigned long size,
935 unsigned long end_pfn)
939 unsigned long ret_paddr;
941 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
943 /* retry over all memory */
945 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
948 panic("numa.c: cannot allocate %lu bytes for node %d",
951 ret = __va(ret_paddr);
954 * We initialize the nodes in numeric order: 0, 1, 2...
955 * and hand over control from the MEMBLOCK allocator to the
956 * bootmem allocator. If this function is called for
957 * node 5, then we know that all nodes <5 are using the
958 * bootmem allocator instead of the MEMBLOCK allocator.
960 * So, check the nid from which this allocation came
961 * and double check to see if we need to use bootmem
962 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
963 * since it would be useless.
965 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
967 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
970 dbg("alloc_bootmem %p %lx\n", ret, size);
973 memset(ret, 0, size);
977 static struct notifier_block ppc64_numa_nb = {
978 .notifier_call = cpu_numa_callback,
979 .priority = 1 /* Must run before sched domains notifier. */
982 static void __init mark_reserved_regions_for_nid(int nid)
984 struct pglist_data *node = NODE_DATA(nid);
985 struct memblock_region *reg;
987 for_each_memblock(reserved, reg) {
988 unsigned long physbase = reg->base;
989 unsigned long size = reg->size;
990 unsigned long start_pfn = physbase >> PAGE_SHIFT;
991 unsigned long end_pfn = PFN_UP(physbase + size);
992 struct node_active_region node_ar;
993 unsigned long node_end_pfn = pgdat_end_pfn(node);
996 * Check to make sure that this memblock.reserved area is
997 * within the bounds of the node that we care about.
998 * Checking the nid of the start and end points is not
999 * sufficient because the reserved area could span the
1002 if (end_pfn <= node->node_start_pfn ||
1003 start_pfn >= node_end_pfn)
1006 get_node_active_region(start_pfn, &node_ar);
1007 while (start_pfn < end_pfn &&
1008 node_ar.start_pfn < node_ar.end_pfn) {
1009 unsigned long reserve_size = size;
1011 * if reserved region extends past active region
1012 * then trim size to active region
1014 if (end_pfn > node_ar.end_pfn)
1015 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
1018 * Only worry about *this* node, others may not
1019 * yet have valid NODE_DATA().
1021 if (node_ar.nid == nid) {
1022 dbg("reserve_bootmem %lx %lx nid=%d\n",
1023 physbase, reserve_size, node_ar.nid);
1024 reserve_bootmem_node(NODE_DATA(node_ar.nid),
1025 physbase, reserve_size,
1029 * if reserved region is contained in the active region
1032 if (end_pfn <= node_ar.end_pfn)
1036 * reserved region extends past the active region
1037 * get next active region that contains this
1040 start_pfn = node_ar.end_pfn;
1041 physbase = start_pfn << PAGE_SHIFT;
1042 size = size - reserve_size;
1043 get_node_active_region(start_pfn, &node_ar);
1049 void __init do_init_bootmem(void)
1054 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1055 max_pfn = max_low_pfn;
1057 if (parse_numa_properties())
1060 dump_numa_memory_topology();
1062 for_each_online_node(nid) {
1063 unsigned long start_pfn, end_pfn;
1064 void *bootmem_vaddr;
1065 unsigned long bootmap_pages;
1067 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1070 * Allocate the node structure node local if possible
1072 * Be careful moving this around, as it relies on all
1073 * previous nodes' bootmem to be initialized and have
1074 * all reserved areas marked.
1076 NODE_DATA(nid) = careful_zallocation(nid,
1077 sizeof(struct pglist_data),
1078 SMP_CACHE_BYTES, end_pfn);
1080 dbg("node %d\n", nid);
1081 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1083 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1084 NODE_DATA(nid)->node_start_pfn = start_pfn;
1085 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1087 if (NODE_DATA(nid)->node_spanned_pages == 0)
1090 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1091 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1093 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1094 bootmem_vaddr = careful_zallocation(nid,
1095 bootmap_pages << PAGE_SHIFT,
1096 PAGE_SIZE, end_pfn);
1098 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1100 init_bootmem_node(NODE_DATA(nid),
1101 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1102 start_pfn, end_pfn);
1104 free_bootmem_with_active_regions(nid, end_pfn);
1106 * Be very careful about moving this around. Future
1107 * calls to careful_zallocation() depend on this getting
1110 mark_reserved_regions_for_nid(nid);
1111 sparse_memory_present_with_active_regions(nid);
1114 init_bootmem_done = 1;
1117 * Now bootmem is initialised we can create the node to cpumask
1118 * lookup tables and setup the cpu callback to populate them.
1120 setup_node_to_cpumask_map();
1122 reset_numa_cpu_lookup_table();
1123 register_cpu_notifier(&ppc64_numa_nb);
1124 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1125 (void *)(unsigned long)boot_cpuid);
1128 void __init paging_init(void)
1130 unsigned long max_zone_pfns[MAX_NR_ZONES];
1131 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1132 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1133 free_area_init_nodes(max_zone_pfns);
1136 static int __init early_numa(char *p)
1141 if (strstr(p, "off"))
1144 if (strstr(p, "debug"))
1147 p = strstr(p, "fake=");
1149 cmdline = p + strlen("fake=");
1153 early_param("numa", early_numa);
1155 #ifdef CONFIG_MEMORY_HOTPLUG
1157 * Find the node associated with a hot added memory section for
1158 * memory represented in the device tree by the property
1159 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1161 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1162 unsigned long scn_addr)
1165 unsigned int drconf_cell_cnt, rc;
1166 unsigned long lmb_size;
1167 struct assoc_arrays aa;
1170 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1171 if (!drconf_cell_cnt)
1174 lmb_size = of_get_lmb_size(memory);
1178 rc = of_get_assoc_arrays(memory, &aa);
1182 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1183 struct of_drconf_cell drmem;
1185 read_drconf_cell(&drmem, &dm);
1187 /* skip this block if it is reserved or not assigned to
1189 if ((drmem.flags & DRCONF_MEM_RESERVED)
1190 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1193 if ((scn_addr < drmem.base_addr)
1194 || (scn_addr >= (drmem.base_addr + lmb_size)))
1197 nid = of_drconf_to_nid_single(&drmem, &aa);
1205 * Find the node associated with a hot added memory section for memory
1206 * represented in the device tree as a node (i.e. memory@XXXX) for
1209 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1211 struct device_node *memory;
1214 for_each_node_by_type(memory, "memory") {
1215 unsigned long start, size;
1217 const __be32 *memcell_buf;
1220 memcell_buf = of_get_property(memory, "reg", &len);
1221 if (!memcell_buf || len <= 0)
1224 /* ranges in cell */
1225 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1228 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1229 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1231 if ((scn_addr < start) || (scn_addr >= (start + size)))
1234 nid = of_node_to_nid_single(memory);
1242 of_node_put(memory);
1248 * Find the node associated with a hot added memory section. Section
1249 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1250 * sections are fully contained within a single MEMBLOCK.
1252 int hot_add_scn_to_nid(unsigned long scn_addr)
1254 struct device_node *memory = NULL;
1257 if (!numa_enabled || (min_common_depth < 0))
1258 return first_online_node;
1260 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1262 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1263 of_node_put(memory);
1265 nid = hot_add_node_scn_to_nid(scn_addr);
1268 if (nid < 0 || !node_online(nid))
1269 nid = first_online_node;
1271 if (NODE_DATA(nid)->node_spanned_pages)
1274 for_each_online_node(nid) {
1275 if (NODE_DATA(nid)->node_spanned_pages) {
1285 static u64 hot_add_drconf_memory_max(void)
1287 struct device_node *memory = NULL;
1288 unsigned int drconf_cell_cnt = 0;
1290 const __be32 *dm = NULL;
1292 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1294 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1295 lmb_size = of_get_lmb_size(memory);
1296 of_node_put(memory);
1298 return lmb_size * drconf_cell_cnt;
1302 * memory_hotplug_max - return max address of memory that may be added
1304 * This is currently only used on systems that support drconfig memory
1307 u64 memory_hotplug_max(void)
1309 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1311 #endif /* CONFIG_MEMORY_HOTPLUG */
1313 /* Virtual Processor Home Node (VPHN) support */
1314 #ifdef CONFIG_PPC_SPLPAR
1315 struct topology_update_data {
1316 struct topology_update_data *next;
1322 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1323 static cpumask_t cpu_associativity_changes_mask;
1324 static int vphn_enabled;
1325 static int prrn_enabled;
1326 static void reset_topology_timer(void);
1329 * Store the current values of the associativity change counters in the
1332 static void setup_cpu_associativity_change_counters(void)
1336 /* The VPHN feature supports a maximum of 8 reference points */
1337 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1339 for_each_possible_cpu(cpu) {
1341 u8 *counts = vphn_cpu_change_counts[cpu];
1342 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1344 for (i = 0; i < distance_ref_points_depth; i++)
1345 counts[i] = hypervisor_counts[i];
1350 * The hypervisor maintains a set of 8 associativity change counters in
1351 * the VPA of each cpu that correspond to the associativity levels in the
1352 * ibm,associativity-reference-points property. When an associativity
1353 * level changes, the corresponding counter is incremented.
1355 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1356 * node associativity levels have changed.
1358 * Returns the number of cpus with unhandled associativity changes.
1360 static int update_cpu_associativity_changes_mask(void)
1363 cpumask_t *changes = &cpu_associativity_changes_mask;
1365 for_each_possible_cpu(cpu) {
1367 u8 *counts = vphn_cpu_change_counts[cpu];
1368 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1370 for (i = 0; i < distance_ref_points_depth; i++) {
1371 if (hypervisor_counts[i] != counts[i]) {
1372 counts[i] = hypervisor_counts[i];
1377 cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1378 cpu = cpu_last_thread_sibling(cpu);
1382 return cpumask_weight(changes);
1386 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1387 * the complete property we have to add the length in the first cell.
1389 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1392 * Convert the associativity domain numbers returned from the hypervisor
1393 * to the sequence they would appear in the ibm,associativity property.
1395 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1397 int i, nr_assoc_doms = 0;
1398 const __be16 *field = (const __be16 *) packed;
1400 #define VPHN_FIELD_UNUSED (0xffff)
1401 #define VPHN_FIELD_MSB (0x8000)
1402 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1404 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1405 if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1406 /* All significant fields processed, and remaining
1407 * fields contain the reserved value of all 1's.
1410 unpacked[i] = *((__be32 *)field);
1412 } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1413 /* Data is in the lower 15 bits of this field */
1414 unpacked[i] = cpu_to_be32(
1415 be16_to_cpup(field) & VPHN_FIELD_MASK);
1419 /* Data is in the lower 15 bits of this field
1420 * concatenated with the next 16 bit field
1422 unpacked[i] = *((__be32 *)field);
1428 /* The first cell contains the length of the property */
1429 unpacked[0] = cpu_to_be32(nr_assoc_doms);
1431 return nr_assoc_doms;
1435 * Retrieve the new associativity information for a virtual processor's
1438 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1441 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1443 int hwcpu = get_hard_smp_processor_id(cpu);
1445 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1446 vphn_unpack_associativity(retbuf, associativity);
1451 static long vphn_get_associativity(unsigned long cpu,
1452 __be32 *associativity)
1456 rc = hcall_vphn(cpu, associativity);
1461 "VPHN is not supported. Disabling polling...\n");
1462 stop_topology_update();
1466 "hcall_vphn() experienced a hardware fault "
1467 "preventing VPHN. Disabling polling...\n");
1468 stop_topology_update();
1475 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1476 * characteristics change. This function doesn't perform any locking and is
1477 * only safe to call from stop_machine().
1479 static int update_cpu_topology(void *data)
1481 struct topology_update_data *update;
1487 cpu = smp_processor_id();
1489 for (update = data; update; update = update->next) {
1490 if (cpu != update->cpu)
1493 unmap_cpu_from_node(update->cpu);
1494 map_cpu_to_node(update->cpu, update->new_nid);
1501 static int update_lookup_table(void *data)
1503 struct topology_update_data *update;
1509 * Upon topology update, the numa-cpu lookup table needs to be updated
1510 * for all threads in the core, including offline CPUs, to ensure that
1511 * future hotplug operations respect the cpu-to-node associativity
1514 for (update = data; update; update = update->next) {
1517 nid = update->new_nid;
1518 base = cpu_first_thread_sibling(update->cpu);
1520 for (j = 0; j < threads_per_core; j++) {
1521 update_numa_cpu_lookup_table(base + j, nid);
1529 * Update the node maps and sysfs entries for each cpu whose home node
1530 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1532 int arch_update_cpu_topology(void)
1534 unsigned int cpu, sibling, changed = 0;
1535 struct topology_update_data *updates, *ud;
1536 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1537 cpumask_t updated_cpus;
1539 int weight, new_nid, i = 0;
1541 weight = cpumask_weight(&cpu_associativity_changes_mask);
1545 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1549 cpumask_clear(&updated_cpus);
1551 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1553 * If siblings aren't flagged for changes, updates list
1554 * will be too short. Skip on this update and set for next
1557 if (!cpumask_subset(cpu_sibling_mask(cpu),
1558 &cpu_associativity_changes_mask)) {
1559 pr_info("Sibling bits not set for associativity "
1560 "change, cpu%d\n", cpu);
1561 cpumask_or(&cpu_associativity_changes_mask,
1562 &cpu_associativity_changes_mask,
1563 cpu_sibling_mask(cpu));
1564 cpu = cpu_last_thread_sibling(cpu);
1568 /* Use associativity from first thread for all siblings */
1569 vphn_get_associativity(cpu, associativity);
1570 new_nid = associativity_to_nid(associativity);
1571 if (new_nid < 0 || !node_online(new_nid))
1572 new_nid = first_online_node;
1574 if (new_nid == numa_cpu_lookup_table[cpu]) {
1575 cpumask_andnot(&cpu_associativity_changes_mask,
1576 &cpu_associativity_changes_mask,
1577 cpu_sibling_mask(cpu));
1578 cpu = cpu_last_thread_sibling(cpu);
1582 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1585 ud->new_nid = new_nid;
1586 ud->old_nid = numa_cpu_lookup_table[sibling];
1587 cpumask_set_cpu(sibling, &updated_cpus);
1589 ud->next = &updates[i];
1591 cpu = cpu_last_thread_sibling(cpu);
1594 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1597 * Update the numa-cpu lookup table with the new mappings, even for
1598 * offline CPUs. It is best to perform this update from the stop-
1601 stop_machine(update_lookup_table, &updates[0],
1602 cpumask_of(raw_smp_processor_id()));
1604 for (ud = &updates[0]; ud; ud = ud->next) {
1605 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1606 register_cpu_under_node(ud->cpu, ud->new_nid);
1608 dev = get_cpu_device(ud->cpu);
1610 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1611 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1619 static void topology_work_fn(struct work_struct *work)
1621 rebuild_sched_domains();
1623 static DECLARE_WORK(topology_work, topology_work_fn);
1625 static void topology_schedule_update(void)
1627 schedule_work(&topology_work);
1630 static void topology_timer_fn(unsigned long ignored)
1632 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1633 topology_schedule_update();
1634 else if (vphn_enabled) {
1635 if (update_cpu_associativity_changes_mask() > 0)
1636 topology_schedule_update();
1637 reset_topology_timer();
1640 static struct timer_list topology_timer =
1641 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1643 static void reset_topology_timer(void)
1645 topology_timer.data = 0;
1646 topology_timer.expires = jiffies + 60 * HZ;
1647 mod_timer(&topology_timer, topology_timer.expires);
1652 static void stage_topology_update(int core_id)
1654 cpumask_or(&cpu_associativity_changes_mask,
1655 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1656 reset_topology_timer();
1659 static int dt_update_callback(struct notifier_block *nb,
1660 unsigned long action, void *data)
1662 struct of_prop_reconfig *update;
1663 int rc = NOTIFY_DONE;
1666 case OF_RECONFIG_UPDATE_PROPERTY:
1667 update = (struct of_prop_reconfig *)data;
1668 if (!of_prop_cmp(update->dn->type, "cpu") &&
1669 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1671 of_property_read_u32(update->dn, "reg", &core_id);
1672 stage_topology_update(core_id);
1681 static struct notifier_block dt_update_nb = {
1682 .notifier_call = dt_update_callback,
1688 * Start polling for associativity changes.
1690 int start_topology_update(void)
1694 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1695 if (!prrn_enabled) {
1699 rc = of_reconfig_notifier_register(&dt_update_nb);
1702 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1703 lppaca_shared_proc(get_lppaca())) {
1704 if (!vphn_enabled) {
1707 setup_cpu_associativity_change_counters();
1708 init_timer_deferrable(&topology_timer);
1709 reset_topology_timer();
1717 * Disable polling for VPHN associativity changes.
1719 int stop_topology_update(void)
1726 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1728 } else if (vphn_enabled) {
1730 rc = del_timer_sync(&topology_timer);
1736 int prrn_is_enabled(void)
1738 return prrn_enabled;
1741 static int topology_read(struct seq_file *file, void *v)
1743 if (vphn_enabled || prrn_enabled)
1744 seq_puts(file, "on\n");
1746 seq_puts(file, "off\n");
1751 static int topology_open(struct inode *inode, struct file *file)
1753 return single_open(file, topology_read, NULL);
1756 static ssize_t topology_write(struct file *file, const char __user *buf,
1757 size_t count, loff_t *off)
1759 char kbuf[4]; /* "on" or "off" plus null. */
1762 read_len = count < 3 ? count : 3;
1763 if (copy_from_user(kbuf, buf, read_len))
1766 kbuf[read_len] = '\0';
1768 if (!strncmp(kbuf, "on", 2))
1769 start_topology_update();
1770 else if (!strncmp(kbuf, "off", 3))
1771 stop_topology_update();
1778 static const struct file_operations topology_ops = {
1780 .write = topology_write,
1781 .open = topology_open,
1782 .release = single_release
1785 static int topology_update_init(void)
1787 start_topology_update();
1788 proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1792 device_initcall(topology_update_init);
1793 #endif /* CONFIG_PPC_SPLPAR */