4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
11 #include <linux/threads.h>
12 #include <linux/bootmem.h>
13 #include <linux/init.h>
15 #include <linux/mmzone.h>
16 #include <linux/module.h>
17 #include <linux/nodemask.h>
18 #include <linux/cpu.h>
19 #include <linux/notifier.h>
20 #include <asm/sparsemem.h>
22 #include <asm/system.h>
25 static int numa_enabled = 1;
27 static int numa_debug;
28 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
30 int numa_cpu_lookup_table[NR_CPUS];
31 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
32 struct pglist_data *node_data[MAX_NUMNODES];
34 EXPORT_SYMBOL(numa_cpu_lookup_table);
35 EXPORT_SYMBOL(numa_cpumask_lookup_table);
36 EXPORT_SYMBOL(node_data);
38 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
39 static int min_common_depth;
40 static int n_mem_addr_cells, n_mem_size_cells;
42 static void __cpuinit map_cpu_to_node(int cpu, int node)
44 numa_cpu_lookup_table[cpu] = node;
46 dbg("adding cpu %d to node %d\n", cpu, node);
48 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
49 cpu_set(cpu, numa_cpumask_lookup_table[node]);
52 #ifdef CONFIG_HOTPLUG_CPU
53 static void unmap_cpu_from_node(unsigned long cpu)
55 int node = numa_cpu_lookup_table[cpu];
57 dbg("removing cpu %lu from node %d\n", cpu, node);
59 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
60 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
62 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
66 #endif /* CONFIG_HOTPLUG_CPU */
68 static struct device_node * __cpuinit find_cpu_node(unsigned int cpu)
70 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
71 struct device_node *cpu_node = NULL;
72 const unsigned int *interrupt_server, *reg;
75 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
76 /* Try interrupt server first */
77 interrupt_server = get_property(cpu_node,
78 "ibm,ppc-interrupt-server#s", &len);
80 len = len / sizeof(u32);
82 if (interrupt_server && (len > 0)) {
84 if (interrupt_server[len] == hw_cpuid)
88 reg = get_property(cpu_node, "reg", &len);
89 if (reg && (len > 0) && (reg[0] == hw_cpuid))
97 /* must hold reference to node during call */
98 static const int *of_get_associativity(struct device_node *dev)
100 return get_property(dev, "ibm,associativity", NULL);
103 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
106 static int of_node_to_nid_single(struct device_node *device)
109 const unsigned int *tmp;
111 if (min_common_depth == -1)
114 tmp = of_get_associativity(device);
118 if (tmp[0] >= min_common_depth)
119 nid = tmp[min_common_depth];
121 /* POWER4 LPAR uses 0xffff as invalid node */
122 if (nid == 0xffff || nid >= MAX_NUMNODES)
128 /* Walk the device tree upwards, looking for an associativity id */
129 int of_node_to_nid(struct device_node *device)
131 struct device_node *tmp;
136 nid = of_node_to_nid_single(device);
141 device = of_get_parent(tmp);
148 EXPORT_SYMBOL_GPL(of_node_to_nid);
151 * In theory, the "ibm,associativity" property may contain multiple
152 * associativity lists because a resource may be multiply connected
153 * into the machine. This resource then has different associativity
154 * characteristics relative to its multiple connections. We ignore
155 * this for now. We also assume that all cpu and memory sets have
156 * their distances represented at a common level. This won't be
157 * true for heirarchical NUMA.
159 * In any case the ibm,associativity-reference-points should give
160 * the correct depth for a normal NUMA system.
162 * - Dave Hansen <haveblue@us.ibm.com>
164 static int __init find_min_common_depth(void)
167 const unsigned int *ref_points;
168 struct device_node *rtas_root;
171 rtas_root = of_find_node_by_path("/rtas");
177 * this property is 2 32-bit integers, each representing a level of
178 * depth in the associativity nodes. The first is for an SMP
179 * configuration (should be all 0's) and the second is for a normal
180 * NUMA configuration.
182 ref_points = get_property(rtas_root,
183 "ibm,associativity-reference-points", &len);
185 if ((len >= 1) && ref_points) {
186 depth = ref_points[1];
188 dbg("NUMA: ibm,associativity-reference-points not found.\n");
191 of_node_put(rtas_root);
196 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
198 struct device_node *memory = NULL;
200 memory = of_find_node_by_type(memory, "memory");
202 panic("numa.c: No memory nodes found!");
204 *n_addr_cells = prom_n_addr_cells(memory);
205 *n_size_cells = prom_n_size_cells(memory);
209 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
211 unsigned long result = 0;
214 result = (result << 32) | **buf;
221 * Figure out to which domain a cpu belongs and stick it there.
222 * Return the id of the domain used.
224 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
227 struct device_node *cpu = find_cpu_node(lcpu);
234 nid = of_node_to_nid_single(cpu);
236 if (nid < 0 || !node_online(nid))
237 nid = any_online_node(NODE_MASK_ALL);
239 map_cpu_to_node(lcpu, nid);
246 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
247 unsigned long action,
250 unsigned long lcpu = (unsigned long)hcpu;
251 int ret = NOTIFY_DONE;
255 numa_setup_cpu(lcpu);
258 #ifdef CONFIG_HOTPLUG_CPU
260 case CPU_UP_CANCELED:
261 unmap_cpu_from_node(lcpu);
270 * Check and possibly modify a memory region to enforce the memory limit.
272 * Returns the size the region should have to enforce the memory limit.
273 * This will either be the original value of size, a truncated value,
274 * or zero. If the returned value of size is 0 the region should be
275 * discarded as it lies wholy above the memory limit.
277 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
281 * We use lmb_end_of_DRAM() in here instead of memory_limit because
282 * we've already adjusted it for the limit and it takes care of
283 * having memory holes below the limit.
289 if (start + size <= lmb_end_of_DRAM())
292 if (start >= lmb_end_of_DRAM())
295 return lmb_end_of_DRAM() - start;
298 static int __init parse_numa_properties(void)
300 struct device_node *cpu = NULL;
301 struct device_node *memory = NULL;
305 if (numa_enabled == 0) {
306 printk(KERN_WARNING "NUMA disabled by user\n");
310 min_common_depth = find_min_common_depth();
312 if (min_common_depth < 0)
313 return min_common_depth;
315 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
318 * Even though we connect cpus to numa domains later in SMP
319 * init, we need to know the node ids now. This is because
320 * each node to be onlined must have NODE_DATA etc backing it.
322 for_each_present_cpu(i) {
325 cpu = find_cpu_node(i);
327 nid = of_node_to_nid_single(cpu);
331 * Don't fall back to default_nid yet -- we will plug
332 * cpus into nodes once the memory scan has discovered
337 node_set_online(nid);
340 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
342 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
347 const unsigned int *memcell_buf;
350 memcell_buf = get_property(memory,
351 "linux,usable-memory", &len);
352 if (!memcell_buf || len <= 0)
353 memcell_buf = get_property(memory, "reg", &len);
354 if (!memcell_buf || len <= 0)
358 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
360 /* these are order-sensitive, and modify the buffer pointer */
361 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
362 size = read_n_cells(n_mem_size_cells, &memcell_buf);
365 * Assumption: either all memory nodes or none will
366 * have associativity properties. If none, then
367 * everything goes to default_nid.
369 nid = of_node_to_nid_single(memory);
372 node_set_online(nid);
374 if (!(size = numa_enforce_memory_limit(start, size))) {
381 add_active_range(nid, start >> PAGE_SHIFT,
382 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
391 static void __init setup_nonnuma(void)
393 unsigned long top_of_ram = lmb_end_of_DRAM();
394 unsigned long total_ram = lmb_phys_mem_size();
395 unsigned long start_pfn, end_pfn;
398 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
399 top_of_ram, total_ram);
400 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
401 (top_of_ram - total_ram) >> 20);
403 for (i = 0; i < lmb.memory.cnt; ++i) {
404 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
405 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
406 add_active_range(0, start_pfn, end_pfn);
411 void __init dump_numa_cpu_topology(void)
414 unsigned int cpu, count;
416 if (min_common_depth == -1 || !numa_enabled)
419 for_each_online_node(node) {
420 printk(KERN_DEBUG "Node %d CPUs:", node);
424 * If we used a CPU iterator here we would miss printing
425 * the holes in the cpumap.
427 for (cpu = 0; cpu < NR_CPUS; cpu++) {
428 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
434 printk("-%u", cpu - 1);
440 printk("-%u", NR_CPUS - 1);
445 static void __init dump_numa_memory_topology(void)
450 if (min_common_depth == -1 || !numa_enabled)
453 for_each_online_node(node) {
456 printk(KERN_DEBUG "Node %d Memory:", node);
460 for (i = 0; i < lmb_end_of_DRAM();
461 i += (1 << SECTION_SIZE_BITS)) {
462 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
480 * Allocate some memory, satisfying the lmb or bootmem allocator where
481 * required. nid is the preferred node and end is the physical address of
482 * the highest address in the node.
484 * Returns the physical address of the memory.
486 static void __init *careful_allocation(int nid, unsigned long size,
488 unsigned long end_pfn)
491 unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
493 /* retry over all memory */
495 ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
498 panic("numa.c: cannot allocate %lu bytes on node %d",
502 * If the memory came from a previously allocated node, we must
503 * retry with the bootmem allocator.
505 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
507 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
511 panic("numa.c: cannot allocate %lu bytes on node %d",
516 dbg("alloc_bootmem %lx %lx\n", ret, size);
522 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
523 .notifier_call = cpu_numa_callback,
524 .priority = 1 /* Must run before sched domains notifier. */
527 void __init do_init_bootmem(void)
533 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
534 max_pfn = max_low_pfn;
536 if (parse_numa_properties())
539 dump_numa_memory_topology();
541 register_cpu_notifier(&ppc64_numa_nb);
542 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
543 (void *)(unsigned long)boot_cpuid);
545 for_each_online_node(nid) {
546 unsigned long start_pfn, end_pfn;
547 unsigned long bootmem_paddr;
548 unsigned long bootmap_pages;
550 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
552 /* Allocate the node structure node local if possible */
553 NODE_DATA(nid) = careful_allocation(nid,
554 sizeof(struct pglist_data),
555 SMP_CACHE_BYTES, end_pfn);
556 NODE_DATA(nid) = __va(NODE_DATA(nid));
557 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
559 dbg("node %d\n", nid);
560 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
562 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
563 NODE_DATA(nid)->node_start_pfn = start_pfn;
564 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
566 if (NODE_DATA(nid)->node_spanned_pages == 0)
569 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
570 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
572 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
573 bootmem_paddr = (unsigned long)careful_allocation(nid,
574 bootmap_pages << PAGE_SHIFT,
576 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
578 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
580 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
583 free_bootmem_with_active_regions(nid, end_pfn);
585 /* Mark reserved regions on this node */
586 for (i = 0; i < lmb.reserved.cnt; i++) {
587 unsigned long physbase = lmb.reserved.region[i].base;
588 unsigned long size = lmb.reserved.region[i].size;
589 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
590 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
592 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
593 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
596 if (physbase < end_paddr &&
597 (physbase+size) > start_paddr) {
599 if (physbase < start_paddr) {
600 size -= start_paddr - physbase;
601 physbase = start_paddr;
604 if (size > end_paddr - physbase)
605 size = end_paddr - physbase;
607 dbg("reserve_bootmem %lx %lx\n", physbase,
609 reserve_bootmem_node(NODE_DATA(nid), physbase,
614 sparse_memory_present_with_active_regions(nid);
618 void __init paging_init(void)
620 unsigned long max_zone_pfns[MAX_NR_ZONES];
621 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
622 max_zone_pfns[ZONE_DMA] = lmb_end_of_DRAM() >> PAGE_SHIFT;
623 free_area_init_nodes(max_zone_pfns);
626 static int __init early_numa(char *p)
631 if (strstr(p, "off"))
634 if (strstr(p, "debug"))
639 early_param("numa", early_numa);
641 #ifdef CONFIG_MEMORY_HOTPLUG
643 * Find the node associated with a hot added memory section. Section
644 * corresponds to a SPARSEMEM section, not an LMB. It is assumed that
645 * sections are fully contained within a single LMB.
647 int hot_add_scn_to_nid(unsigned long scn_addr)
649 struct device_node *memory = NULL;
651 int default_nid = any_online_node(NODE_MASK_ALL);
654 if (!numa_enabled || (min_common_depth < 0))
657 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
658 unsigned long start, size;
660 const unsigned int *memcell_buf;
663 memcell_buf = get_property(memory, "reg", &len);
664 if (!memcell_buf || len <= 0)
668 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
670 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
671 size = read_n_cells(n_mem_size_cells, &memcell_buf);
672 nid = of_node_to_nid_single(memory);
674 /* Domains not present at boot default to 0 */
675 if (nid < 0 || !node_online(nid))
678 if ((scn_addr >= start) && (scn_addr < (start + size))) {
683 if (--ranges) /* process all ranges in cell */
686 BUG(); /* section address should be found above */
689 /* Temporary code to ensure that returned node is not empty */
692 while (NODE_DATA(nid)->node_spanned_pages == 0) {
693 node_clear(nid, nodes);
694 nid = any_online_node(nodes);
698 #endif /* CONFIG_MEMORY_HOTPLUG */