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 <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)
66 unsigned int node, num = 0;
68 /* setup nr_node_ids if not done yet */
69 if (nr_node_ids == MAX_NUMNODES) {
70 for_each_node_mask(node, node_possible_map)
72 nr_node_ids = num + 1;
75 /* allocate the map */
76 for (node = 0; node < nr_node_ids; node++)
77 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
79 /* cpumask_of_node() will now work */
80 dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
83 static int __init fake_numa_create_new_node(unsigned long end_pfn,
86 unsigned long long mem;
88 static unsigned int fake_nid;
89 static unsigned long long curr_boundary;
92 * Modify node id, iff we started creating NUMA nodes
93 * We want to continue from where we left of the last time
98 * In case there are no more arguments to parse, the
99 * node_id should be the same as the last fake node id
100 * (we've handled this above).
105 mem = memparse(p, &p);
109 if (mem < curr_boundary)
114 if ((end_pfn << PAGE_SHIFT) > mem) {
116 * Skip commas and spaces
118 while (*p == ',' || *p == ' ' || *p == '\t')
124 dbg("created new fake_node with id %d\n", fake_nid);
131 * get_node_active_region - Return active region containing pfn
132 * Active range returned is empty if none found.
133 * @pfn: The page to return the region for
134 * @node_ar: Returned set to the active region containing @pfn
136 static void __init get_node_active_region(unsigned long pfn,
137 struct node_active_region *node_ar)
139 unsigned long start_pfn, end_pfn;
142 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
143 if (pfn >= start_pfn && pfn < end_pfn) {
145 node_ar->start_pfn = start_pfn;
146 node_ar->end_pfn = end_pfn;
152 static void map_cpu_to_node(int cpu, int node)
154 numa_cpu_lookup_table[cpu] = node;
156 dbg("adding cpu %d to node %d\n", cpu, node);
158 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
159 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
162 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
163 static void unmap_cpu_from_node(unsigned long cpu)
165 int node = numa_cpu_lookup_table[cpu];
167 dbg("removing cpu %lu from node %d\n", cpu, node);
169 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
170 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
172 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
176 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
178 /* must hold reference to node during call */
179 static const int *of_get_associativity(struct device_node *dev)
181 return of_get_property(dev, "ibm,associativity", NULL);
185 * Returns the property linux,drconf-usable-memory if
186 * it exists (the property exists only in kexec/kdump kernels,
187 * added by kexec-tools)
189 static const u32 *of_get_usable_memory(struct device_node *memory)
193 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
194 if (!prop || len < sizeof(unsigned int))
199 int __node_distance(int a, int b)
202 int distance = LOCAL_DISTANCE;
205 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
207 for (i = 0; i < distance_ref_points_depth; i++) {
208 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
211 /* Double the distance for each NUMA level */
218 static void initialize_distance_lookup_table(int nid,
219 const unsigned int *associativity)
226 for (i = 0; i < distance_ref_points_depth; i++) {
227 distance_lookup_table[nid][i] =
228 associativity[distance_ref_points[i]];
232 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
235 static int associativity_to_nid(const unsigned int *associativity)
239 if (min_common_depth == -1)
242 if (associativity[0] >= min_common_depth)
243 nid = associativity[min_common_depth];
245 /* POWER4 LPAR uses 0xffff as invalid node */
246 if (nid == 0xffff || nid >= MAX_NUMNODES)
249 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
250 initialize_distance_lookup_table(nid, associativity);
256 /* Returns the nid associated with the given device tree node,
257 * or -1 if not found.
259 static int of_node_to_nid_single(struct device_node *device)
262 const unsigned int *tmp;
264 tmp = of_get_associativity(device);
266 nid = associativity_to_nid(tmp);
270 /* Walk the device tree upwards, looking for an associativity id */
271 int of_node_to_nid(struct device_node *device)
273 struct device_node *tmp;
278 nid = of_node_to_nid_single(device);
283 device = of_get_parent(tmp);
290 EXPORT_SYMBOL_GPL(of_node_to_nid);
292 static int __init find_min_common_depth(void)
295 struct device_node *root;
297 if (firmware_has_feature(FW_FEATURE_OPAL))
298 root = of_find_node_by_path("/ibm,opal");
300 root = of_find_node_by_path("/rtas");
302 root = of_find_node_by_path("/");
305 * This property is a set of 32-bit integers, each representing
306 * an index into the ibm,associativity nodes.
308 * With form 0 affinity the first integer is for an SMP configuration
309 * (should be all 0's) and the second is for a normal NUMA
310 * configuration. We have only one level of NUMA.
312 * With form 1 affinity the first integer is the most significant
313 * NUMA boundary and the following are progressively less significant
314 * boundaries. There can be more than one level of NUMA.
316 distance_ref_points = of_get_property(root,
317 "ibm,associativity-reference-points",
318 &distance_ref_points_depth);
320 if (!distance_ref_points) {
321 dbg("NUMA: ibm,associativity-reference-points not found.\n");
325 distance_ref_points_depth /= sizeof(int);
327 if (firmware_has_feature(FW_FEATURE_OPAL) ||
328 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
329 dbg("Using form 1 affinity\n");
333 if (form1_affinity) {
334 depth = distance_ref_points[0];
336 if (distance_ref_points_depth < 2) {
337 printk(KERN_WARNING "NUMA: "
338 "short ibm,associativity-reference-points\n");
342 depth = distance_ref_points[1];
346 * Warn and cap if the hardware supports more than
347 * MAX_DISTANCE_REF_POINTS domains.
349 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
350 printk(KERN_WARNING "NUMA: distance array capped at "
351 "%d entries\n", MAX_DISTANCE_REF_POINTS);
352 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
363 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
365 struct device_node *memory = NULL;
367 memory = of_find_node_by_type(memory, "memory");
369 panic("numa.c: No memory nodes found!");
371 *n_addr_cells = of_n_addr_cells(memory);
372 *n_size_cells = of_n_size_cells(memory);
376 static unsigned long read_n_cells(int n, const unsigned int **buf)
378 unsigned long result = 0;
381 result = (result << 32) | **buf;
388 * Read the next memblock list entry from the ibm,dynamic-memory property
389 * and return the information in the provided of_drconf_cell structure.
391 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
395 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
398 drmem->drc_index = cp[0];
399 drmem->reserved = cp[1];
400 drmem->aa_index = cp[2];
401 drmem->flags = cp[3];
407 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
409 * The layout of the ibm,dynamic-memory property is a number N of memblock
410 * list entries followed by N memblock list entries. Each memblock list entry
411 * contains information as laid out in the of_drconf_cell struct above.
413 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
418 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
419 if (!prop || len < sizeof(unsigned int))
424 /* Now that we know the number of entries, revalidate the size
425 * of the property read in to ensure we have everything
427 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
435 * Retrieve and validate the ibm,lmb-size property for drconf memory
436 * from the device tree.
438 static u64 of_get_lmb_size(struct device_node *memory)
443 prop = of_get_property(memory, "ibm,lmb-size", &len);
444 if (!prop || len < sizeof(unsigned int))
447 return read_n_cells(n_mem_size_cells, &prop);
450 struct assoc_arrays {
457 * Retrieve and validate the list of associativity arrays for drconf
458 * memory from the ibm,associativity-lookup-arrays property of the
461 * The layout of the ibm,associativity-lookup-arrays property is a number N
462 * indicating the number of associativity arrays, followed by a number M
463 * indicating the size of each associativity array, followed by a list
464 * of N associativity arrays.
466 static int of_get_assoc_arrays(struct device_node *memory,
467 struct assoc_arrays *aa)
472 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
473 if (!prop || len < 2 * sizeof(unsigned int))
476 aa->n_arrays = *prop++;
477 aa->array_sz = *prop++;
479 /* Now that we know the number of arrays and size of each array,
480 * revalidate the size of the property read in.
482 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
490 * This is like of_node_to_nid_single() for memory represented in the
491 * ibm,dynamic-reconfiguration-memory node.
493 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
494 struct assoc_arrays *aa)
497 int nid = default_nid;
500 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
501 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
502 drmem->aa_index < aa->n_arrays) {
503 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
504 nid = aa->arrays[index];
506 if (nid == 0xffff || nid >= MAX_NUMNODES)
514 * Figure out to which domain a cpu belongs and stick it there.
515 * Return the id of the domain used.
517 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
520 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
527 nid = of_node_to_nid_single(cpu);
529 if (nid < 0 || !node_online(nid))
530 nid = first_online_node;
532 map_cpu_to_node(lcpu, nid);
539 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
540 unsigned long action,
543 unsigned long lcpu = (unsigned long)hcpu;
544 int ret = NOTIFY_DONE;
548 case CPU_UP_PREPARE_FROZEN:
549 numa_setup_cpu(lcpu);
552 #ifdef CONFIG_HOTPLUG_CPU
554 case CPU_DEAD_FROZEN:
555 case CPU_UP_CANCELED:
556 case CPU_UP_CANCELED_FROZEN:
557 unmap_cpu_from_node(lcpu);
566 * Check and possibly modify a memory region to enforce the memory limit.
568 * Returns the size the region should have to enforce the memory limit.
569 * This will either be the original value of size, a truncated value,
570 * or zero. If the returned value of size is 0 the region should be
571 * discarded as it lies wholly above the memory limit.
573 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
577 * We use memblock_end_of_DRAM() in here instead of memory_limit because
578 * we've already adjusted it for the limit and it takes care of
579 * having memory holes below the limit. Also, in the case of
580 * iommu_is_off, memory_limit is not set but is implicitly enforced.
583 if (start + size <= memblock_end_of_DRAM())
586 if (start >= memblock_end_of_DRAM())
589 return memblock_end_of_DRAM() - start;
593 * Reads the counter for a given entry in
594 * linux,drconf-usable-memory property
596 static inline int __init read_usm_ranges(const u32 **usm)
599 * For each lmb in ibm,dynamic-memory a corresponding
600 * entry in linux,drconf-usable-memory property contains
601 * a counter followed by that many (base, size) duple.
602 * read the counter from linux,drconf-usable-memory
604 return read_n_cells(n_mem_size_cells, usm);
608 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
609 * node. This assumes n_mem_{addr,size}_cells have been set.
611 static void __init parse_drconf_memory(struct device_node *memory)
613 const u32 *uninitialized_var(dm), *usm;
614 unsigned int n, rc, ranges, is_kexec_kdump = 0;
615 unsigned long lmb_size, base, size, sz;
617 struct assoc_arrays aa = { .arrays = NULL };
619 n = of_get_drconf_memory(memory, &dm);
623 lmb_size = of_get_lmb_size(memory);
627 rc = of_get_assoc_arrays(memory, &aa);
631 /* check if this is a kexec/kdump kernel */
632 usm = of_get_usable_memory(memory);
636 for (; n != 0; --n) {
637 struct of_drconf_cell drmem;
639 read_drconf_cell(&drmem, &dm);
641 /* skip this block if the reserved bit is set in flags (0x80)
642 or if the block is not assigned to this partition (0x8) */
643 if ((drmem.flags & DRCONF_MEM_RESERVED)
644 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
647 base = drmem.base_addr;
651 if (is_kexec_kdump) {
652 ranges = read_usm_ranges(&usm);
653 if (!ranges) /* there are no (base, size) duple */
657 if (is_kexec_kdump) {
658 base = read_n_cells(n_mem_addr_cells, &usm);
659 size = read_n_cells(n_mem_size_cells, &usm);
661 nid = of_drconf_to_nid_single(&drmem, &aa);
662 fake_numa_create_new_node(
663 ((base + size) >> PAGE_SHIFT),
665 node_set_online(nid);
666 sz = numa_enforce_memory_limit(base, size);
668 memblock_set_node(base, sz, nid);
673 static int __init parse_numa_properties(void)
675 struct device_node *memory;
679 if (numa_enabled == 0) {
680 printk(KERN_WARNING "NUMA disabled by user\n");
684 min_common_depth = find_min_common_depth();
686 if (min_common_depth < 0)
687 return min_common_depth;
689 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
692 * Even though we connect cpus to numa domains later in SMP
693 * init, we need to know the node ids now. This is because
694 * each node to be onlined must have NODE_DATA etc backing it.
696 for_each_present_cpu(i) {
697 struct device_node *cpu;
700 cpu = of_get_cpu_node(i, NULL);
702 nid = of_node_to_nid_single(cpu);
706 * Don't fall back to default_nid yet -- we will plug
707 * cpus into nodes once the memory scan has discovered
712 node_set_online(nid);
715 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
717 for_each_node_by_type(memory, "memory") {
722 const unsigned int *memcell_buf;
725 memcell_buf = of_get_property(memory,
726 "linux,usable-memory", &len);
727 if (!memcell_buf || len <= 0)
728 memcell_buf = of_get_property(memory, "reg", &len);
729 if (!memcell_buf || len <= 0)
733 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
735 /* these are order-sensitive, and modify the buffer pointer */
736 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
737 size = read_n_cells(n_mem_size_cells, &memcell_buf);
740 * Assumption: either all memory nodes or none will
741 * have associativity properties. If none, then
742 * everything goes to default_nid.
744 nid = of_node_to_nid_single(memory);
748 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
749 node_set_online(nid);
751 if (!(size = numa_enforce_memory_limit(start, size))) {
758 memblock_set_node(start, size, nid);
765 * Now do the same thing for each MEMBLOCK listed in the
766 * ibm,dynamic-memory property in the
767 * ibm,dynamic-reconfiguration-memory node.
769 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
771 parse_drconf_memory(memory);
776 static void __init setup_nonnuma(void)
778 unsigned long top_of_ram = memblock_end_of_DRAM();
779 unsigned long total_ram = memblock_phys_mem_size();
780 unsigned long start_pfn, end_pfn;
781 unsigned int nid = 0;
782 struct memblock_region *reg;
784 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
785 top_of_ram, total_ram);
786 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
787 (top_of_ram - total_ram) >> 20);
789 for_each_memblock(memory, reg) {
790 start_pfn = memblock_region_memory_base_pfn(reg);
791 end_pfn = memblock_region_memory_end_pfn(reg);
793 fake_numa_create_new_node(end_pfn, &nid);
794 memblock_set_node(PFN_PHYS(start_pfn),
795 PFN_PHYS(end_pfn - start_pfn), nid);
796 node_set_online(nid);
800 void __init dump_numa_cpu_topology(void)
803 unsigned int cpu, count;
805 if (min_common_depth == -1 || !numa_enabled)
808 for_each_online_node(node) {
809 printk(KERN_DEBUG "Node %d CPUs:", node);
813 * If we used a CPU iterator here we would miss printing
814 * the holes in the cpumap.
816 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
817 if (cpumask_test_cpu(cpu,
818 node_to_cpumask_map[node])) {
824 printk("-%u", cpu - 1);
830 printk("-%u", nr_cpu_ids - 1);
835 static void __init dump_numa_memory_topology(void)
840 if (min_common_depth == -1 || !numa_enabled)
843 for_each_online_node(node) {
846 printk(KERN_DEBUG "Node %d Memory:", node);
850 for (i = 0; i < memblock_end_of_DRAM();
851 i += (1 << SECTION_SIZE_BITS)) {
852 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
870 * Allocate some memory, satisfying the memblock or bootmem allocator where
871 * required. nid is the preferred node and end is the physical address of
872 * the highest address in the node.
874 * Returns the virtual address of the memory.
876 static void __init *careful_zallocation(int nid, unsigned long size,
878 unsigned long end_pfn)
882 unsigned long ret_paddr;
884 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
886 /* retry over all memory */
888 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
891 panic("numa.c: cannot allocate %lu bytes for node %d",
894 ret = __va(ret_paddr);
897 * We initialize the nodes in numeric order: 0, 1, 2...
898 * and hand over control from the MEMBLOCK allocator to the
899 * bootmem allocator. If this function is called for
900 * node 5, then we know that all nodes <5 are using the
901 * bootmem allocator instead of the MEMBLOCK allocator.
903 * So, check the nid from which this allocation came
904 * and double check to see if we need to use bootmem
905 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
906 * since it would be useless.
908 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
910 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
913 dbg("alloc_bootmem %p %lx\n", ret, size);
916 memset(ret, 0, size);
920 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
921 .notifier_call = cpu_numa_callback,
922 .priority = 1 /* Must run before sched domains notifier. */
925 static void __init mark_reserved_regions_for_nid(int nid)
927 struct pglist_data *node = NODE_DATA(nid);
928 struct memblock_region *reg;
930 for_each_memblock(reserved, reg) {
931 unsigned long physbase = reg->base;
932 unsigned long size = reg->size;
933 unsigned long start_pfn = physbase >> PAGE_SHIFT;
934 unsigned long end_pfn = PFN_UP(physbase + size);
935 struct node_active_region node_ar;
936 unsigned long node_end_pfn = node->node_start_pfn +
937 node->node_spanned_pages;
940 * Check to make sure that this memblock.reserved area is
941 * within the bounds of the node that we care about.
942 * Checking the nid of the start and end points is not
943 * sufficient because the reserved area could span the
946 if (end_pfn <= node->node_start_pfn ||
947 start_pfn >= node_end_pfn)
950 get_node_active_region(start_pfn, &node_ar);
951 while (start_pfn < end_pfn &&
952 node_ar.start_pfn < node_ar.end_pfn) {
953 unsigned long reserve_size = size;
955 * if reserved region extends past active region
956 * then trim size to active region
958 if (end_pfn > node_ar.end_pfn)
959 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
962 * Only worry about *this* node, others may not
963 * yet have valid NODE_DATA().
965 if (node_ar.nid == nid) {
966 dbg("reserve_bootmem %lx %lx nid=%d\n",
967 physbase, reserve_size, node_ar.nid);
968 reserve_bootmem_node(NODE_DATA(node_ar.nid),
969 physbase, reserve_size,
973 * if reserved region is contained in the active region
976 if (end_pfn <= node_ar.end_pfn)
980 * reserved region extends past the active region
981 * get next active region that contains this
984 start_pfn = node_ar.end_pfn;
985 physbase = start_pfn << PAGE_SHIFT;
986 size = size - reserve_size;
987 get_node_active_region(start_pfn, &node_ar);
993 void __init do_init_bootmem(void)
998 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
999 max_pfn = max_low_pfn;
1001 if (parse_numa_properties())
1004 dump_numa_memory_topology();
1006 for_each_online_node(nid) {
1007 unsigned long start_pfn, end_pfn;
1008 void *bootmem_vaddr;
1009 unsigned long bootmap_pages;
1011 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1014 * Allocate the node structure node local if possible
1016 * Be careful moving this around, as it relies on all
1017 * previous nodes' bootmem to be initialized and have
1018 * all reserved areas marked.
1020 NODE_DATA(nid) = careful_zallocation(nid,
1021 sizeof(struct pglist_data),
1022 SMP_CACHE_BYTES, end_pfn);
1024 dbg("node %d\n", nid);
1025 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1027 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1028 NODE_DATA(nid)->node_start_pfn = start_pfn;
1029 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1031 if (NODE_DATA(nid)->node_spanned_pages == 0)
1034 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1035 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1037 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1038 bootmem_vaddr = careful_zallocation(nid,
1039 bootmap_pages << PAGE_SHIFT,
1040 PAGE_SIZE, end_pfn);
1042 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1044 init_bootmem_node(NODE_DATA(nid),
1045 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1046 start_pfn, end_pfn);
1048 free_bootmem_with_active_regions(nid, end_pfn);
1050 * Be very careful about moving this around. Future
1051 * calls to careful_zallocation() depend on this getting
1054 mark_reserved_regions_for_nid(nid);
1055 sparse_memory_present_with_active_regions(nid);
1058 init_bootmem_done = 1;
1061 * Now bootmem is initialised we can create the node to cpumask
1062 * lookup tables and setup the cpu callback to populate them.
1064 setup_node_to_cpumask_map();
1066 register_cpu_notifier(&ppc64_numa_nb);
1067 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1068 (void *)(unsigned long)boot_cpuid);
1071 void __init paging_init(void)
1073 unsigned long max_zone_pfns[MAX_NR_ZONES];
1074 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1075 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1076 free_area_init_nodes(max_zone_pfns);
1079 static int __init early_numa(char *p)
1084 if (strstr(p, "off"))
1087 if (strstr(p, "debug"))
1090 p = strstr(p, "fake=");
1092 cmdline = p + strlen("fake=");
1096 early_param("numa", early_numa);
1098 #ifdef CONFIG_MEMORY_HOTPLUG
1100 * Find the node associated with a hot added memory section for
1101 * memory represented in the device tree by the property
1102 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1104 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1105 unsigned long scn_addr)
1108 unsigned int drconf_cell_cnt, rc;
1109 unsigned long lmb_size;
1110 struct assoc_arrays aa;
1113 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1114 if (!drconf_cell_cnt)
1117 lmb_size = of_get_lmb_size(memory);
1121 rc = of_get_assoc_arrays(memory, &aa);
1125 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1126 struct of_drconf_cell drmem;
1128 read_drconf_cell(&drmem, &dm);
1130 /* skip this block if it is reserved or not assigned to
1132 if ((drmem.flags & DRCONF_MEM_RESERVED)
1133 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1136 if ((scn_addr < drmem.base_addr)
1137 || (scn_addr >= (drmem.base_addr + lmb_size)))
1140 nid = of_drconf_to_nid_single(&drmem, &aa);
1148 * Find the node associated with a hot added memory section for memory
1149 * represented in the device tree as a node (i.e. memory@XXXX) for
1152 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1154 struct device_node *memory;
1157 for_each_node_by_type(memory, "memory") {
1158 unsigned long start, size;
1160 const unsigned int *memcell_buf;
1163 memcell_buf = of_get_property(memory, "reg", &len);
1164 if (!memcell_buf || len <= 0)
1167 /* ranges in cell */
1168 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1171 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1172 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1174 if ((scn_addr < start) || (scn_addr >= (start + size)))
1177 nid = of_node_to_nid_single(memory);
1185 of_node_put(memory);
1191 * Find the node associated with a hot added memory section. Section
1192 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1193 * sections are fully contained within a single MEMBLOCK.
1195 int hot_add_scn_to_nid(unsigned long scn_addr)
1197 struct device_node *memory = NULL;
1200 if (!numa_enabled || (min_common_depth < 0))
1201 return first_online_node;
1203 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1205 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1206 of_node_put(memory);
1208 nid = hot_add_node_scn_to_nid(scn_addr);
1211 if (nid < 0 || !node_online(nid))
1212 nid = first_online_node;
1214 if (NODE_DATA(nid)->node_spanned_pages)
1217 for_each_online_node(nid) {
1218 if (NODE_DATA(nid)->node_spanned_pages) {
1228 static u64 hot_add_drconf_memory_max(void)
1230 struct device_node *memory = NULL;
1231 unsigned int drconf_cell_cnt = 0;
1235 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1237 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1238 lmb_size = of_get_lmb_size(memory);
1239 of_node_put(memory);
1241 return lmb_size * drconf_cell_cnt;
1245 * memory_hotplug_max - return max address of memory that may be added
1247 * This is currently only used on systems that support drconfig memory
1250 u64 memory_hotplug_max(void)
1252 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1254 #endif /* CONFIG_MEMORY_HOTPLUG */
1256 /* Virtual Processor Home Node (VPHN) support */
1257 #ifdef CONFIG_PPC_SPLPAR
1258 struct topology_update_data {
1259 struct topology_update_data *next;
1265 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1266 static cpumask_t cpu_associativity_changes_mask;
1267 static int vphn_enabled;
1268 static int prrn_enabled;
1269 static void reset_topology_timer(void);
1272 * Store the current values of the associativity change counters in the
1275 static void setup_cpu_associativity_change_counters(void)
1279 /* The VPHN feature supports a maximum of 8 reference points */
1280 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1282 for_each_possible_cpu(cpu) {
1284 u8 *counts = vphn_cpu_change_counts[cpu];
1285 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1287 for (i = 0; i < distance_ref_points_depth; i++)
1288 counts[i] = hypervisor_counts[i];
1293 * The hypervisor maintains a set of 8 associativity change counters in
1294 * the VPA of each cpu that correspond to the associativity levels in the
1295 * ibm,associativity-reference-points property. When an associativity
1296 * level changes, the corresponding counter is incremented.
1298 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1299 * node associativity levels have changed.
1301 * Returns the number of cpus with unhandled associativity changes.
1303 static int update_cpu_associativity_changes_mask(void)
1306 cpumask_t *changes = &cpu_associativity_changes_mask;
1308 for_each_possible_cpu(cpu) {
1310 u8 *counts = vphn_cpu_change_counts[cpu];
1311 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1313 for (i = 0; i < distance_ref_points_depth; i++) {
1314 if (hypervisor_counts[i] != counts[i]) {
1315 counts[i] = hypervisor_counts[i];
1320 cpumask_set_cpu(cpu, changes);
1324 return cpumask_weight(changes);
1328 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1329 * the complete property we have to add the length in the first cell.
1331 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1334 * Convert the associativity domain numbers returned from the hypervisor
1335 * to the sequence they would appear in the ibm,associativity property.
1337 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1339 int i, nr_assoc_doms = 0;
1340 const u16 *field = (const u16*) packed;
1342 #define VPHN_FIELD_UNUSED (0xffff)
1343 #define VPHN_FIELD_MSB (0x8000)
1344 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1346 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1347 if (*field == VPHN_FIELD_UNUSED) {
1348 /* All significant fields processed, and remaining
1349 * fields contain the reserved value of all 1's.
1352 unpacked[i] = *((u32*)field);
1354 } else if (*field & VPHN_FIELD_MSB) {
1355 /* Data is in the lower 15 bits of this field */
1356 unpacked[i] = *field & VPHN_FIELD_MASK;
1360 /* Data is in the lower 15 bits of this field
1361 * concatenated with the next 16 bit field
1363 unpacked[i] = *((u32*)field);
1369 /* The first cell contains the length of the property */
1370 unpacked[0] = nr_assoc_doms;
1372 return nr_assoc_doms;
1376 * Retrieve the new associativity information for a virtual processor's
1379 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1382 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1384 int hwcpu = get_hard_smp_processor_id(cpu);
1386 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1387 vphn_unpack_associativity(retbuf, associativity);
1392 static long vphn_get_associativity(unsigned long cpu,
1393 unsigned int *associativity)
1397 rc = hcall_vphn(cpu, associativity);
1402 "VPHN is not supported. Disabling polling...\n");
1403 stop_topology_update();
1407 "hcall_vphn() experienced a hardware fault "
1408 "preventing VPHN. Disabling polling...\n");
1409 stop_topology_update();
1416 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1417 * characteristics change. This function doesn't perform any locking and is
1418 * only safe to call from stop_machine().
1420 static int update_cpu_topology(void *data)
1422 struct topology_update_data *update;
1430 for (update = data; update; update = update->next) {
1431 if (cpu != update->cpu)
1434 unregister_cpu_under_node(update->cpu, update->old_nid);
1435 unmap_cpu_from_node(update->cpu);
1436 map_cpu_to_node(update->cpu, update->new_nid);
1437 register_cpu_under_node(update->cpu, update->new_nid);
1444 * Update the node maps and sysfs entries for each cpu whose home node
1445 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1447 int arch_update_cpu_topology(void)
1449 unsigned int cpu, changed = 0;
1450 struct topology_update_data *updates, *ud;
1451 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1455 weight = cpumask_weight(&cpu_associativity_changes_mask);
1459 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1463 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1466 vphn_get_associativity(cpu, associativity);
1467 ud->new_nid = associativity_to_nid(associativity);
1469 if (ud->new_nid < 0 || !node_online(ud->new_nid))
1470 ud->new_nid = first_online_node;
1472 ud->old_nid = numa_cpu_lookup_table[cpu];
1475 ud->next = &updates[i];
1478 stop_machine(update_cpu_topology, &updates[0], cpu_online_mask);
1480 for (ud = &updates[0]; ud; ud = ud->next) {
1481 dev = get_cpu_device(ud->cpu);
1483 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1484 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1492 static void topology_work_fn(struct work_struct *work)
1494 rebuild_sched_domains();
1496 static DECLARE_WORK(topology_work, topology_work_fn);
1498 void topology_schedule_update(void)
1500 schedule_work(&topology_work);
1503 static void topology_timer_fn(unsigned long ignored)
1505 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1506 topology_schedule_update();
1507 else if (vphn_enabled) {
1508 if (update_cpu_associativity_changes_mask() > 0)
1509 topology_schedule_update();
1510 reset_topology_timer();
1513 static struct timer_list topology_timer =
1514 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1516 static void reset_topology_timer(void)
1518 topology_timer.data = 0;
1519 topology_timer.expires = jiffies + 60 * HZ;
1520 mod_timer(&topology_timer, topology_timer.expires);
1523 static void stage_topology_update(int core_id)
1525 cpumask_or(&cpu_associativity_changes_mask,
1526 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1527 reset_topology_timer();
1530 static int dt_update_callback(struct notifier_block *nb,
1531 unsigned long action, void *data)
1533 struct of_prop_reconfig *update;
1534 int rc = NOTIFY_DONE;
1537 case OF_RECONFIG_UPDATE_PROPERTY:
1538 update = (struct of_prop_reconfig *)data;
1539 if (!of_prop_cmp(update->dn->type, "cpu") &&
1540 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1542 of_property_read_u32(update->dn, "reg", &core_id);
1543 stage_topology_update(core_id);
1552 static struct notifier_block dt_update_nb = {
1553 .notifier_call = dt_update_callback,
1557 * Start polling for associativity changes.
1559 int start_topology_update(void)
1563 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1564 if (!prrn_enabled) {
1567 rc = of_reconfig_notifier_register(&dt_update_nb);
1569 } else if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1570 get_lppaca()->shared_proc) {
1571 /* Disabled until races with load balancing are fixed */
1572 if (!vphn_enabled) {
1575 setup_cpu_associativity_change_counters();
1576 init_timer_deferrable(&topology_timer);
1577 reset_topology_timer();
1583 __initcall(start_topology_update);
1586 * Disable polling for VPHN associativity changes.
1588 int stop_topology_update(void)
1594 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1595 } else if (vphn_enabled) {
1597 rc = del_timer_sync(&topology_timer);
1602 #endif /* CONFIG_PPC_SPLPAR */