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/sparsemem.h>
33 #include <asm/firmware.h>
35 #include <asm/hvcall.h>
36 #include <asm/setup.h>
39 static int numa_enabled = 1;
41 static char *cmdline __initdata;
43 static int numa_debug;
44 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
46 int numa_cpu_lookup_table[NR_CPUS];
47 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
48 struct pglist_data *node_data[MAX_NUMNODES];
50 EXPORT_SYMBOL(numa_cpu_lookup_table);
51 EXPORT_SYMBOL(node_to_cpumask_map);
52 EXPORT_SYMBOL(node_data);
54 static int min_common_depth;
55 static int n_mem_addr_cells, n_mem_size_cells;
56 static int form1_affinity;
58 #define MAX_DISTANCE_REF_POINTS 4
59 static int distance_ref_points_depth;
60 static const unsigned int *distance_ref_points;
61 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
64 * Allocate node_to_cpumask_map based on number of available nodes
65 * Requires node_possible_map to be valid.
67 * Note: cpumask_of_node() is not valid until after this is done.
69 static void __init setup_node_to_cpumask_map(void)
71 unsigned int node, num = 0;
73 /* setup nr_node_ids if not done yet */
74 if (nr_node_ids == MAX_NUMNODES) {
75 for_each_node_mask(node, node_possible_map)
77 nr_node_ids = num + 1;
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 map_cpu_to_node(int cpu, int node)
159 numa_cpu_lookup_table[cpu] = node;
161 dbg("adding cpu %d to node %d\n", cpu, node);
163 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
164 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
167 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
168 static void unmap_cpu_from_node(unsigned long cpu)
170 int node = numa_cpu_lookup_table[cpu];
172 dbg("removing cpu %lu from node %d\n", cpu, node);
174 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
175 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
177 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
181 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
183 /* must hold reference to node during call */
184 static const int *of_get_associativity(struct device_node *dev)
186 return of_get_property(dev, "ibm,associativity", NULL);
190 * Returns the property linux,drconf-usable-memory if
191 * it exists (the property exists only in kexec/kdump kernels,
192 * added by kexec-tools)
194 static const u32 *of_get_usable_memory(struct device_node *memory)
198 prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
199 if (!prop || len < sizeof(unsigned int))
204 int __node_distance(int a, int b)
207 int distance = LOCAL_DISTANCE;
210 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
212 for (i = 0; i < distance_ref_points_depth; i++) {
213 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
216 /* Double the distance for each NUMA level */
223 static void initialize_distance_lookup_table(int nid,
224 const unsigned int *associativity)
231 for (i = 0; i < distance_ref_points_depth; i++) {
232 distance_lookup_table[nid][i] =
233 associativity[distance_ref_points[i]];
237 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
240 static int associativity_to_nid(const unsigned int *associativity)
244 if (min_common_depth == -1)
247 if (associativity[0] >= min_common_depth)
248 nid = associativity[min_common_depth];
250 /* POWER4 LPAR uses 0xffff as invalid node */
251 if (nid == 0xffff || nid >= MAX_NUMNODES)
254 if (nid > 0 && associativity[0] >= distance_ref_points_depth)
255 initialize_distance_lookup_table(nid, associativity);
261 /* Returns the nid associated with the given device tree node,
262 * or -1 if not found.
264 static int of_node_to_nid_single(struct device_node *device)
267 const unsigned int *tmp;
269 tmp = of_get_associativity(device);
271 nid = associativity_to_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 *root;
302 if (firmware_has_feature(FW_FEATURE_OPAL))
303 root = of_find_node_by_path("/ibm,opal");
305 root = of_find_node_by_path("/rtas");
307 root = of_find_node_by_path("/");
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(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 if (firmware_has_feature(FW_FEATURE_OPAL) ||
333 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
334 dbg("Using form 1 affinity\n");
338 if (form1_affinity) {
339 depth = distance_ref_points[0];
341 if (distance_ref_points_depth < 2) {
342 printk(KERN_WARNING "NUMA: "
343 "short ibm,associativity-reference-points\n");
347 depth = distance_ref_points[1];
351 * Warn and cap if the hardware supports more than
352 * MAX_DISTANCE_REF_POINTS domains.
354 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
355 printk(KERN_WARNING "NUMA: distance array capped at "
356 "%d entries\n", MAX_DISTANCE_REF_POINTS);
357 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
368 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
370 struct device_node *memory = NULL;
372 memory = of_find_node_by_type(memory, "memory");
374 panic("numa.c: No memory nodes found!");
376 *n_addr_cells = of_n_addr_cells(memory);
377 *n_size_cells = of_n_size_cells(memory);
381 static unsigned long read_n_cells(int n, const unsigned int **buf)
383 unsigned long result = 0;
386 result = (result << 32) | **buf;
393 * Read the next memblock list entry from the ibm,dynamic-memory property
394 * and return the information in the provided of_drconf_cell structure.
396 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
400 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
403 drmem->drc_index = cp[0];
404 drmem->reserved = cp[1];
405 drmem->aa_index = cp[2];
406 drmem->flags = cp[3];
412 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
414 * The layout of the ibm,dynamic-memory property is a number N of memblock
415 * list entries followed by N memblock list entries. Each memblock list entry
416 * contains information as laid out in the of_drconf_cell struct above.
418 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
423 prop = of_get_property(memory, "ibm,dynamic-memory", &len);
424 if (!prop || len < sizeof(unsigned int))
429 /* Now that we know the number of entries, revalidate the size
430 * of the property read in to ensure we have everything
432 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
440 * Retrieve and validate the ibm,lmb-size property for drconf memory
441 * from the device tree.
443 static u64 of_get_lmb_size(struct device_node *memory)
448 prop = of_get_property(memory, "ibm,lmb-size", &len);
449 if (!prop || len < sizeof(unsigned int))
452 return read_n_cells(n_mem_size_cells, &prop);
455 struct assoc_arrays {
462 * Retrieve and validate the list of associativity arrays for drconf
463 * memory from the ibm,associativity-lookup-arrays property of the
466 * The layout of the ibm,associativity-lookup-arrays property is a number N
467 * indicating the number of associativity arrays, followed by a number M
468 * indicating the size of each associativity array, followed by a list
469 * of N associativity arrays.
471 static int of_get_assoc_arrays(struct device_node *memory,
472 struct assoc_arrays *aa)
477 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
478 if (!prop || len < 2 * sizeof(unsigned int))
481 aa->n_arrays = *prop++;
482 aa->array_sz = *prop++;
484 /* Now that we know the number of arrays and size of each array,
485 * revalidate the size of the property read in.
487 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
495 * This is like of_node_to_nid_single() for memory represented in the
496 * ibm,dynamic-reconfiguration-memory node.
498 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
499 struct assoc_arrays *aa)
502 int nid = default_nid;
505 if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
506 !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
507 drmem->aa_index < aa->n_arrays) {
508 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
509 nid = aa->arrays[index];
511 if (nid == 0xffff || nid >= MAX_NUMNODES)
519 * Figure out to which domain a cpu belongs and stick it there.
520 * Return the id of the domain used.
522 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
525 struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
532 nid = of_node_to_nid_single(cpu);
534 if (nid < 0 || !node_online(nid))
535 nid = first_online_node;
537 map_cpu_to_node(lcpu, nid);
544 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
545 unsigned long action,
548 unsigned long lcpu = (unsigned long)hcpu;
549 int ret = NOTIFY_DONE;
553 case CPU_UP_PREPARE_FROZEN:
554 numa_setup_cpu(lcpu);
557 #ifdef CONFIG_HOTPLUG_CPU
559 case CPU_DEAD_FROZEN:
560 case CPU_UP_CANCELED:
561 case CPU_UP_CANCELED_FROZEN:
562 unmap_cpu_from_node(lcpu);
571 * Check and possibly modify a memory region to enforce the memory limit.
573 * Returns the size the region should have to enforce the memory limit.
574 * This will either be the original value of size, a truncated value,
575 * or zero. If the returned value of size is 0 the region should be
576 * discarded as it lies wholly above the memory limit.
578 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
582 * We use memblock_end_of_DRAM() in here instead of memory_limit because
583 * we've already adjusted it for the limit and it takes care of
584 * having memory holes below the limit. Also, in the case of
585 * iommu_is_off, memory_limit is not set but is implicitly enforced.
588 if (start + size <= memblock_end_of_DRAM())
591 if (start >= memblock_end_of_DRAM())
594 return memblock_end_of_DRAM() - start;
598 * Reads the counter for a given entry in
599 * linux,drconf-usable-memory property
601 static inline int __init read_usm_ranges(const u32 **usm)
604 * For each lmb in ibm,dynamic-memory a corresponding
605 * entry in linux,drconf-usable-memory property contains
606 * a counter followed by that many (base, size) duple.
607 * read the counter from linux,drconf-usable-memory
609 return read_n_cells(n_mem_size_cells, usm);
613 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
614 * node. This assumes n_mem_{addr,size}_cells have been set.
616 static void __init parse_drconf_memory(struct device_node *memory)
618 const u32 *uninitialized_var(dm), *usm;
619 unsigned int n, rc, ranges, is_kexec_kdump = 0;
620 unsigned long lmb_size, base, size, sz;
622 struct assoc_arrays aa = { .arrays = NULL };
624 n = of_get_drconf_memory(memory, &dm);
628 lmb_size = of_get_lmb_size(memory);
632 rc = of_get_assoc_arrays(memory, &aa);
636 /* check if this is a kexec/kdump kernel */
637 usm = of_get_usable_memory(memory);
641 for (; n != 0; --n) {
642 struct of_drconf_cell drmem;
644 read_drconf_cell(&drmem, &dm);
646 /* skip this block if the reserved bit is set in flags (0x80)
647 or if the block is not assigned to this partition (0x8) */
648 if ((drmem.flags & DRCONF_MEM_RESERVED)
649 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
652 base = drmem.base_addr;
656 if (is_kexec_kdump) {
657 ranges = read_usm_ranges(&usm);
658 if (!ranges) /* there are no (base, size) duple */
662 if (is_kexec_kdump) {
663 base = read_n_cells(n_mem_addr_cells, &usm);
664 size = read_n_cells(n_mem_size_cells, &usm);
666 nid = of_drconf_to_nid_single(&drmem, &aa);
667 fake_numa_create_new_node(
668 ((base + size) >> PAGE_SHIFT),
670 node_set_online(nid);
671 sz = numa_enforce_memory_limit(base, size);
673 memblock_set_node(base, sz, nid);
678 static int __init parse_numa_properties(void)
680 struct device_node *memory;
684 if (numa_enabled == 0) {
685 printk(KERN_WARNING "NUMA disabled by user\n");
689 min_common_depth = find_min_common_depth();
691 if (min_common_depth < 0)
692 return min_common_depth;
694 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
697 * Even though we connect cpus to numa domains later in SMP
698 * init, we need to know the node ids now. This is because
699 * each node to be onlined must have NODE_DATA etc backing it.
701 for_each_present_cpu(i) {
702 struct device_node *cpu;
705 cpu = of_get_cpu_node(i, NULL);
707 nid = of_node_to_nid_single(cpu);
711 * Don't fall back to default_nid yet -- we will plug
712 * cpus into nodes once the memory scan has discovered
717 node_set_online(nid);
720 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
722 for_each_node_by_type(memory, "memory") {
727 const unsigned int *memcell_buf;
730 memcell_buf = of_get_property(memory,
731 "linux,usable-memory", &len);
732 if (!memcell_buf || len <= 0)
733 memcell_buf = of_get_property(memory, "reg", &len);
734 if (!memcell_buf || len <= 0)
738 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
740 /* these are order-sensitive, and modify the buffer pointer */
741 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
742 size = read_n_cells(n_mem_size_cells, &memcell_buf);
745 * Assumption: either all memory nodes or none will
746 * have associativity properties. If none, then
747 * everything goes to default_nid.
749 nid = of_node_to_nid_single(memory);
753 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
754 node_set_online(nid);
756 if (!(size = numa_enforce_memory_limit(start, size))) {
763 memblock_set_node(start, size, nid);
770 * Now do the same thing for each MEMBLOCK listed in the
771 * ibm,dynamic-memory property in the
772 * ibm,dynamic-reconfiguration-memory node.
774 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
776 parse_drconf_memory(memory);
781 static void __init setup_nonnuma(void)
783 unsigned long top_of_ram = memblock_end_of_DRAM();
784 unsigned long total_ram = memblock_phys_mem_size();
785 unsigned long start_pfn, end_pfn;
786 unsigned int nid = 0;
787 struct memblock_region *reg;
789 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
790 top_of_ram, total_ram);
791 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
792 (top_of_ram - total_ram) >> 20);
794 for_each_memblock(memory, reg) {
795 start_pfn = memblock_region_memory_base_pfn(reg);
796 end_pfn = memblock_region_memory_end_pfn(reg);
798 fake_numa_create_new_node(end_pfn, &nid);
799 memblock_set_node(PFN_PHYS(start_pfn),
800 PFN_PHYS(end_pfn - start_pfn), nid);
801 node_set_online(nid);
805 void __init dump_numa_cpu_topology(void)
808 unsigned int cpu, count;
810 if (min_common_depth == -1 || !numa_enabled)
813 for_each_online_node(node) {
814 printk(KERN_DEBUG "Node %d CPUs:", node);
818 * If we used a CPU iterator here we would miss printing
819 * the holes in the cpumap.
821 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
822 if (cpumask_test_cpu(cpu,
823 node_to_cpumask_map[node])) {
829 printk("-%u", cpu - 1);
835 printk("-%u", nr_cpu_ids - 1);
840 static void __init dump_numa_memory_topology(void)
845 if (min_common_depth == -1 || !numa_enabled)
848 for_each_online_node(node) {
851 printk(KERN_DEBUG "Node %d Memory:", node);
855 for (i = 0; i < memblock_end_of_DRAM();
856 i += (1 << SECTION_SIZE_BITS)) {
857 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
875 * Allocate some memory, satisfying the memblock or bootmem allocator where
876 * required. nid is the preferred node and end is the physical address of
877 * the highest address in the node.
879 * Returns the virtual address of the memory.
881 static void __init *careful_zallocation(int nid, unsigned long size,
883 unsigned long end_pfn)
887 unsigned long ret_paddr;
889 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
891 /* retry over all memory */
893 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
896 panic("numa.c: cannot allocate %lu bytes for node %d",
899 ret = __va(ret_paddr);
902 * We initialize the nodes in numeric order: 0, 1, 2...
903 * and hand over control from the MEMBLOCK allocator to the
904 * bootmem allocator. If this function is called for
905 * node 5, then we know that all nodes <5 are using the
906 * bootmem allocator instead of the MEMBLOCK allocator.
908 * So, check the nid from which this allocation came
909 * and double check to see if we need to use bootmem
910 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory
911 * since it would be useless.
913 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
915 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
918 dbg("alloc_bootmem %p %lx\n", ret, size);
921 memset(ret, 0, size);
925 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
926 .notifier_call = cpu_numa_callback,
927 .priority = 1 /* Must run before sched domains notifier. */
930 static void __init mark_reserved_regions_for_nid(int nid)
932 struct pglist_data *node = NODE_DATA(nid);
933 struct memblock_region *reg;
935 for_each_memblock(reserved, reg) {
936 unsigned long physbase = reg->base;
937 unsigned long size = reg->size;
938 unsigned long start_pfn = physbase >> PAGE_SHIFT;
939 unsigned long end_pfn = PFN_UP(physbase + size);
940 struct node_active_region node_ar;
941 unsigned long node_end_pfn = node->node_start_pfn +
942 node->node_spanned_pages;
945 * Check to make sure that this memblock.reserved area is
946 * within the bounds of the node that we care about.
947 * Checking the nid of the start and end points is not
948 * sufficient because the reserved area could span the
951 if (end_pfn <= node->node_start_pfn ||
952 start_pfn >= node_end_pfn)
955 get_node_active_region(start_pfn, &node_ar);
956 while (start_pfn < end_pfn &&
957 node_ar.start_pfn < node_ar.end_pfn) {
958 unsigned long reserve_size = size;
960 * if reserved region extends past active region
961 * then trim size to active region
963 if (end_pfn > node_ar.end_pfn)
964 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
967 * Only worry about *this* node, others may not
968 * yet have valid NODE_DATA().
970 if (node_ar.nid == nid) {
971 dbg("reserve_bootmem %lx %lx nid=%d\n",
972 physbase, reserve_size, node_ar.nid);
973 reserve_bootmem_node(NODE_DATA(node_ar.nid),
974 physbase, reserve_size,
978 * if reserved region is contained in the active region
981 if (end_pfn <= node_ar.end_pfn)
985 * reserved region extends past the active region
986 * get next active region that contains this
989 start_pfn = node_ar.end_pfn;
990 physbase = start_pfn << PAGE_SHIFT;
991 size = size - reserve_size;
992 get_node_active_region(start_pfn, &node_ar);
998 void __init do_init_bootmem(void)
1003 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1004 max_pfn = max_low_pfn;
1006 if (parse_numa_properties())
1009 dump_numa_memory_topology();
1011 for_each_online_node(nid) {
1012 unsigned long start_pfn, end_pfn;
1013 void *bootmem_vaddr;
1014 unsigned long bootmap_pages;
1016 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1019 * Allocate the node structure node local if possible
1021 * Be careful moving this around, as it relies on all
1022 * previous nodes' bootmem to be initialized and have
1023 * all reserved areas marked.
1025 NODE_DATA(nid) = careful_zallocation(nid,
1026 sizeof(struct pglist_data),
1027 SMP_CACHE_BYTES, end_pfn);
1029 dbg("node %d\n", nid);
1030 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1032 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1033 NODE_DATA(nid)->node_start_pfn = start_pfn;
1034 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1036 if (NODE_DATA(nid)->node_spanned_pages == 0)
1039 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1040 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1042 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1043 bootmem_vaddr = careful_zallocation(nid,
1044 bootmap_pages << PAGE_SHIFT,
1045 PAGE_SIZE, end_pfn);
1047 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1049 init_bootmem_node(NODE_DATA(nid),
1050 __pa(bootmem_vaddr) >> PAGE_SHIFT,
1051 start_pfn, end_pfn);
1053 free_bootmem_with_active_regions(nid, end_pfn);
1055 * Be very careful about moving this around. Future
1056 * calls to careful_zallocation() depend on this getting
1059 mark_reserved_regions_for_nid(nid);
1060 sparse_memory_present_with_active_regions(nid);
1063 init_bootmem_done = 1;
1066 * Now bootmem is initialised we can create the node to cpumask
1067 * lookup tables and setup the cpu callback to populate them.
1069 setup_node_to_cpumask_map();
1071 register_cpu_notifier(&ppc64_numa_nb);
1072 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1073 (void *)(unsigned long)boot_cpuid);
1076 void __init paging_init(void)
1078 unsigned long max_zone_pfns[MAX_NR_ZONES];
1079 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1080 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1081 free_area_init_nodes(max_zone_pfns);
1084 static int __init early_numa(char *p)
1089 if (strstr(p, "off"))
1092 if (strstr(p, "debug"))
1095 p = strstr(p, "fake=");
1097 cmdline = p + strlen("fake=");
1101 early_param("numa", early_numa);
1103 #ifdef CONFIG_MEMORY_HOTPLUG
1105 * Find the node associated with a hot added memory section for
1106 * memory represented in the device tree by the property
1107 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1109 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1110 unsigned long scn_addr)
1113 unsigned int drconf_cell_cnt, rc;
1114 unsigned long lmb_size;
1115 struct assoc_arrays aa;
1118 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1119 if (!drconf_cell_cnt)
1122 lmb_size = of_get_lmb_size(memory);
1126 rc = of_get_assoc_arrays(memory, &aa);
1130 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1131 struct of_drconf_cell drmem;
1133 read_drconf_cell(&drmem, &dm);
1135 /* skip this block if it is reserved or not assigned to
1137 if ((drmem.flags & DRCONF_MEM_RESERVED)
1138 || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1141 if ((scn_addr < drmem.base_addr)
1142 || (scn_addr >= (drmem.base_addr + lmb_size)))
1145 nid = of_drconf_to_nid_single(&drmem, &aa);
1153 * Find the node associated with a hot added memory section for memory
1154 * represented in the device tree as a node (i.e. memory@XXXX) for
1157 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1159 struct device_node *memory;
1162 for_each_node_by_type(memory, "memory") {
1163 unsigned long start, size;
1165 const unsigned int *memcell_buf;
1168 memcell_buf = of_get_property(memory, "reg", &len);
1169 if (!memcell_buf || len <= 0)
1172 /* ranges in cell */
1173 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1176 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1177 size = read_n_cells(n_mem_size_cells, &memcell_buf);
1179 if ((scn_addr < start) || (scn_addr >= (start + size)))
1182 nid = of_node_to_nid_single(memory);
1190 of_node_put(memory);
1196 * Find the node associated with a hot added memory section. Section
1197 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that
1198 * sections are fully contained within a single MEMBLOCK.
1200 int hot_add_scn_to_nid(unsigned long scn_addr)
1202 struct device_node *memory = NULL;
1205 if (!numa_enabled || (min_common_depth < 0))
1206 return first_online_node;
1208 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1210 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1211 of_node_put(memory);
1213 nid = hot_add_node_scn_to_nid(scn_addr);
1216 if (nid < 0 || !node_online(nid))
1217 nid = first_online_node;
1219 if (NODE_DATA(nid)->node_spanned_pages)
1222 for_each_online_node(nid) {
1223 if (NODE_DATA(nid)->node_spanned_pages) {
1233 static u64 hot_add_drconf_memory_max(void)
1235 struct device_node *memory = NULL;
1236 unsigned int drconf_cell_cnt = 0;
1240 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1242 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1243 lmb_size = of_get_lmb_size(memory);
1244 of_node_put(memory);
1246 return lmb_size * drconf_cell_cnt;
1250 * memory_hotplug_max - return max address of memory that may be added
1252 * This is currently only used on systems that support drconfig memory
1255 u64 memory_hotplug_max(void)
1257 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1259 #endif /* CONFIG_MEMORY_HOTPLUG */
1261 /* Virtual Processor Home Node (VPHN) support */
1262 #ifdef CONFIG_PPC_SPLPAR
1263 struct topology_update_data {
1264 struct topology_update_data *next;
1270 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1271 static cpumask_t cpu_associativity_changes_mask;
1272 static int vphn_enabled;
1273 static int prrn_enabled;
1274 static void reset_topology_timer(void);
1277 * Store the current values of the associativity change counters in the
1280 static void setup_cpu_associativity_change_counters(void)
1284 /* The VPHN feature supports a maximum of 8 reference points */
1285 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1287 for_each_possible_cpu(cpu) {
1289 u8 *counts = vphn_cpu_change_counts[cpu];
1290 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1292 for (i = 0; i < distance_ref_points_depth; i++)
1293 counts[i] = hypervisor_counts[i];
1298 * The hypervisor maintains a set of 8 associativity change counters in
1299 * the VPA of each cpu that correspond to the associativity levels in the
1300 * ibm,associativity-reference-points property. When an associativity
1301 * level changes, the corresponding counter is incremented.
1303 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1304 * node associativity levels have changed.
1306 * Returns the number of cpus with unhandled associativity changes.
1308 static int update_cpu_associativity_changes_mask(void)
1311 cpumask_t *changes = &cpu_associativity_changes_mask;
1313 for_each_possible_cpu(cpu) {
1315 u8 *counts = vphn_cpu_change_counts[cpu];
1316 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1318 for (i = 0; i < distance_ref_points_depth; i++) {
1319 if (hypervisor_counts[i] != counts[i]) {
1320 counts[i] = hypervisor_counts[i];
1325 cpumask_set_cpu(cpu, changes);
1329 return cpumask_weight(changes);
1333 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1334 * the complete property we have to add the length in the first cell.
1336 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1339 * Convert the associativity domain numbers returned from the hypervisor
1340 * to the sequence they would appear in the ibm,associativity property.
1342 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1344 int i, nr_assoc_doms = 0;
1345 const u16 *field = (const u16*) packed;
1347 #define VPHN_FIELD_UNUSED (0xffff)
1348 #define VPHN_FIELD_MSB (0x8000)
1349 #define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
1351 for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1352 if (*field == VPHN_FIELD_UNUSED) {
1353 /* All significant fields processed, and remaining
1354 * fields contain the reserved value of all 1's.
1357 unpacked[i] = *((u32*)field);
1359 } else if (*field & VPHN_FIELD_MSB) {
1360 /* Data is in the lower 15 bits of this field */
1361 unpacked[i] = *field & VPHN_FIELD_MASK;
1365 /* Data is in the lower 15 bits of this field
1366 * concatenated with the next 16 bit field
1368 unpacked[i] = *((u32*)field);
1374 /* The first cell contains the length of the property */
1375 unpacked[0] = nr_assoc_doms;
1377 return nr_assoc_doms;
1381 * Retrieve the new associativity information for a virtual processor's
1384 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1387 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1389 int hwcpu = get_hard_smp_processor_id(cpu);
1391 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1392 vphn_unpack_associativity(retbuf, associativity);
1397 static long vphn_get_associativity(unsigned long cpu,
1398 unsigned int *associativity)
1402 rc = hcall_vphn(cpu, associativity);
1407 "VPHN is not supported. Disabling polling...\n");
1408 stop_topology_update();
1412 "hcall_vphn() experienced a hardware fault "
1413 "preventing VPHN. Disabling polling...\n");
1414 stop_topology_update();
1421 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1422 * characteristics change. This function doesn't perform any locking and is
1423 * only safe to call from stop_machine().
1425 static int update_cpu_topology(void *data)
1427 struct topology_update_data *update;
1435 for (update = data; update; update = update->next) {
1436 if (cpu != update->cpu)
1439 unregister_cpu_under_node(update->cpu, update->old_nid);
1440 unmap_cpu_from_node(update->cpu);
1441 map_cpu_to_node(update->cpu, update->new_nid);
1443 register_cpu_under_node(update->cpu, update->new_nid);
1450 * Update the node maps and sysfs entries for each cpu whose home node
1451 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1453 int arch_update_cpu_topology(void)
1455 unsigned int cpu, changed = 0;
1456 struct topology_update_data *updates, *ud;
1457 unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1458 cpumask_t updated_cpus;
1462 weight = cpumask_weight(&cpu_associativity_changes_mask);
1466 updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1470 cpumask_clear(&updated_cpus);
1472 for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1475 vphn_get_associativity(cpu, associativity);
1476 ud->new_nid = associativity_to_nid(associativity);
1478 if (ud->new_nid < 0 || !node_online(ud->new_nid))
1479 ud->new_nid = first_online_node;
1481 ud->old_nid = numa_cpu_lookup_table[cpu];
1482 cpumask_set_cpu(cpu, &updated_cpus);
1485 ud->next = &updates[i];
1488 stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1490 for (ud = &updates[0]; ud; ud = ud->next) {
1491 dev = get_cpu_device(ud->cpu);
1493 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1494 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1502 static void topology_work_fn(struct work_struct *work)
1504 rebuild_sched_domains();
1506 static DECLARE_WORK(topology_work, topology_work_fn);
1508 void topology_schedule_update(void)
1510 schedule_work(&topology_work);
1513 static void topology_timer_fn(unsigned long ignored)
1515 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1516 topology_schedule_update();
1517 else if (vphn_enabled) {
1518 if (update_cpu_associativity_changes_mask() > 0)
1519 topology_schedule_update();
1520 reset_topology_timer();
1523 static struct timer_list topology_timer =
1524 TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1526 static void reset_topology_timer(void)
1528 topology_timer.data = 0;
1529 topology_timer.expires = jiffies + 60 * HZ;
1530 mod_timer(&topology_timer, topology_timer.expires);
1533 static void stage_topology_update(int core_id)
1535 cpumask_or(&cpu_associativity_changes_mask,
1536 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1537 reset_topology_timer();
1540 static int dt_update_callback(struct notifier_block *nb,
1541 unsigned long action, void *data)
1543 struct of_prop_reconfig *update;
1544 int rc = NOTIFY_DONE;
1547 case OF_RECONFIG_UPDATE_PROPERTY:
1548 update = (struct of_prop_reconfig *)data;
1549 if (!of_prop_cmp(update->dn->type, "cpu") &&
1550 !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1552 of_property_read_u32(update->dn, "reg", &core_id);
1553 stage_topology_update(core_id);
1562 static struct notifier_block dt_update_nb = {
1563 .notifier_call = dt_update_callback,
1567 * Start polling for associativity changes.
1569 int start_topology_update(void)
1573 if (firmware_has_feature(FW_FEATURE_PRRN)) {
1574 if (!prrn_enabled) {
1577 rc = of_reconfig_notifier_register(&dt_update_nb);
1579 } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1580 get_lppaca()->shared_proc) {
1581 if (!vphn_enabled) {
1584 setup_cpu_associativity_change_counters();
1585 init_timer_deferrable(&topology_timer);
1586 reset_topology_timer();
1594 * Disable polling for VPHN associativity changes.
1596 int stop_topology_update(void)
1602 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1603 } else if (vphn_enabled) {
1605 rc = del_timer_sync(&topology_timer);
1611 int prrn_is_enabled(void)
1613 return prrn_enabled;
1616 static int topology_read(struct seq_file *file, void *v)
1618 if (vphn_enabled || prrn_enabled)
1619 seq_puts(file, "on\n");
1621 seq_puts(file, "off\n");
1626 static int topology_open(struct inode *inode, struct file *file)
1628 return single_open(file, topology_read, NULL);
1631 static ssize_t topology_write(struct file *file, const char __user *buf,
1632 size_t count, loff_t *off)
1634 char kbuf[4]; /* "on" or "off" plus null. */
1637 read_len = count < 3 ? count : 3;
1638 if (copy_from_user(kbuf, buf, read_len))
1641 kbuf[read_len] = '\0';
1643 if (!strncmp(kbuf, "on", 2))
1644 start_topology_update();
1645 else if (!strncmp(kbuf, "off", 3))
1646 stop_topology_update();
1653 static const struct file_operations topology_ops = {
1655 .write = topology_write,
1656 .open = topology_open,
1657 .release = single_release
1660 static int topology_update_init(void)
1662 start_topology_update();
1663 proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1667 device_initcall(topology_update_init);
1668 #endif /* CONFIG_PPC_SPLPAR */