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>
21 #include <asm/machdep.h>
22 #include <asm/abs_addr.h>
24 static int numa_enabled = 1;
26 static int numa_debug;
27 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
30 #define ARRAY_INITIALISER -1
32 #define ARRAY_INITIALISER 0
35 int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
37 char *numa_memory_lookup_table;
38 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
39 int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
41 struct pglist_data *node_data[MAX_NUMNODES];
42 bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
43 static int min_common_depth;
46 * We need somewhere to store start/span for each node until we have
47 * allocated the real node_data structures.
50 unsigned long node_start_pfn;
51 unsigned long node_end_pfn;
52 unsigned long node_present_pages;
53 } init_node_data[MAX_NUMNODES] __initdata;
55 EXPORT_SYMBOL(node_data);
56 EXPORT_SYMBOL(numa_cpu_lookup_table);
57 EXPORT_SYMBOL(numa_memory_lookup_table);
58 EXPORT_SYMBOL(numa_cpumask_lookup_table);
59 EXPORT_SYMBOL(nr_cpus_in_node);
61 static inline void map_cpu_to_node(int cpu, int node)
63 numa_cpu_lookup_table[cpu] = node;
64 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node]))) {
65 cpu_set(cpu, numa_cpumask_lookup_table[node]);
66 nr_cpus_in_node[node]++;
70 #ifdef CONFIG_HOTPLUG_CPU
71 static void unmap_cpu_from_node(unsigned long cpu)
73 int node = numa_cpu_lookup_table[cpu];
75 dbg("removing cpu %lu from node %d\n", cpu, node);
77 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
78 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
79 nr_cpus_in_node[node]--;
81 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
85 #endif /* CONFIG_HOTPLUG_CPU */
87 static struct device_node * __devinit find_cpu_node(unsigned int cpu)
89 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
90 struct device_node *cpu_node = NULL;
91 unsigned int *interrupt_server, *reg;
94 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
95 /* Try interrupt server first */
96 interrupt_server = (unsigned int *)get_property(cpu_node,
97 "ibm,ppc-interrupt-server#s", &len);
99 len = len / sizeof(u32);
101 if (interrupt_server && (len > 0)) {
103 if (interrupt_server[len] == hw_cpuid)
107 reg = (unsigned int *)get_property(cpu_node,
109 if (reg && (len > 0) && (reg[0] == hw_cpuid))
117 /* must hold reference to node during call */
118 static int *of_get_associativity(struct device_node *dev)
120 return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
123 static int of_node_numa_domain(struct device_node *device)
128 if (min_common_depth == -1)
131 tmp = of_get_associativity(device);
132 if (tmp && (tmp[0] >= min_common_depth)) {
133 numa_domain = tmp[min_common_depth];
135 dbg("WARNING: no NUMA information for %s\n",
143 * In theory, the "ibm,associativity" property may contain multiple
144 * associativity lists because a resource may be multiply connected
145 * into the machine. This resource then has different associativity
146 * characteristics relative to its multiple connections. We ignore
147 * this for now. We also assume that all cpu and memory sets have
148 * their distances represented at a common level. This won't be
149 * true for heirarchical NUMA.
151 * In any case the ibm,associativity-reference-points should give
152 * the correct depth for a normal NUMA system.
154 * - Dave Hansen <haveblue@us.ibm.com>
156 static int __init find_min_common_depth(void)
159 unsigned int *ref_points;
160 struct device_node *rtas_root;
163 rtas_root = of_find_node_by_path("/rtas");
169 * this property is 2 32-bit integers, each representing a level of
170 * depth in the associativity nodes. The first is for an SMP
171 * configuration (should be all 0's) and the second is for a normal
172 * NUMA configuration.
174 ref_points = (unsigned int *)get_property(rtas_root,
175 "ibm,associativity-reference-points", &len);
177 if ((len >= 1) && ref_points) {
178 depth = ref_points[1];
180 dbg("WARNING: could not find NUMA "
181 "associativity reference point\n");
184 of_node_put(rtas_root);
189 static int __init get_mem_addr_cells(void)
191 struct device_node *memory = NULL;
194 memory = of_find_node_by_type(memory, "memory");
196 return 0; /* it won't matter */
198 rc = prom_n_addr_cells(memory);
202 static int __init get_mem_size_cells(void)
204 struct device_node *memory = NULL;
207 memory = of_find_node_by_type(memory, "memory");
209 return 0; /* it won't matter */
210 rc = prom_n_size_cells(memory);
214 static unsigned long read_n_cells(int n, unsigned int **buf)
216 unsigned long result = 0;
219 result = (result << 32) | **buf;
226 * Figure out to which domain a cpu belongs and stick it there.
227 * Return the id of the domain used.
229 static int numa_setup_cpu(unsigned long lcpu)
232 struct device_node *cpu = find_cpu_node(lcpu);
239 numa_domain = of_node_numa_domain(cpu);
241 if (numa_domain >= num_online_nodes()) {
243 * POWER4 LPAR uses 0xffff as invalid node,
244 * dont warn in this case.
246 if (numa_domain != 0xffff)
247 printk(KERN_ERR "WARNING: cpu %ld "
248 "maps to invalid NUMA node %d\n",
253 node_set_online(numa_domain);
255 map_cpu_to_node(lcpu, numa_domain);
262 static int cpu_numa_callback(struct notifier_block *nfb,
263 unsigned long action,
266 unsigned long lcpu = (unsigned long)hcpu;
267 int ret = NOTIFY_DONE;
271 if (min_common_depth == -1 || !numa_enabled)
272 map_cpu_to_node(lcpu, 0);
274 numa_setup_cpu(lcpu);
277 #ifdef CONFIG_HOTPLUG_CPU
279 case CPU_UP_CANCELED:
280 unmap_cpu_from_node(lcpu);
289 * Check and possibly modify a memory region to enforce the memory limit.
291 * Returns the size the region should have to enforce the memory limit.
292 * This will either be the original value of size, a truncated value,
293 * or zero. If the returned value of size is 0 the region should be
294 * discarded as it lies wholy above the memory limit.
296 static unsigned long __init numa_enforce_memory_limit(unsigned long start, unsigned long size)
299 * We use lmb_end_of_DRAM() in here instead of memory_limit because
300 * we've already adjusted it for the limit and it takes care of
301 * having memory holes below the limit.
303 extern unsigned long memory_limit;
308 if (start + size <= lmb_end_of_DRAM())
311 if (start >= lmb_end_of_DRAM())
314 return lmb_end_of_DRAM() - start;
317 static int __init parse_numa_properties(void)
319 struct device_node *cpu = NULL;
320 struct device_node *memory = NULL;
321 int addr_cells, size_cells;
323 long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
326 if (numa_enabled == 0) {
327 printk(KERN_WARNING "NUMA disabled by user\n");
331 numa_memory_lookup_table =
332 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
333 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
335 for (i = 0; i < entries ; i++)
336 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
338 min_common_depth = find_min_common_depth();
340 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
341 if (min_common_depth < 0)
342 return min_common_depth;
344 max_domain = numa_setup_cpu(boot_cpuid);
347 * Even though we connect cpus to numa domains later in SMP init,
348 * we need to know the maximum node id now. This is because each
349 * node id must have NODE_DATA etc backing it.
350 * As a result of hotplug we could still have cpus appear later on
351 * with larger node ids. In that case we force the cpu into node 0.
356 cpu = find_cpu_node(i);
359 numa_domain = of_node_numa_domain(cpu);
362 if (numa_domain < MAX_NUMNODES &&
363 max_domain < numa_domain)
364 max_domain = numa_domain;
368 addr_cells = get_mem_addr_cells();
369 size_cells = get_mem_size_cells();
371 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
376 unsigned int *memcell_buf;
379 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
380 if (!memcell_buf || len <= 0)
383 ranges = memory->n_addrs;
385 /* these are order-sensitive, and modify the buffer pointer */
386 start = read_n_cells(addr_cells, &memcell_buf);
387 size = read_n_cells(size_cells, &memcell_buf);
389 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
390 size = _ALIGN_UP(size, MEMORY_INCREMENT);
392 numa_domain = of_node_numa_domain(memory);
394 if (numa_domain >= MAX_NUMNODES) {
395 if (numa_domain != 0xffff)
396 printk(KERN_ERR "WARNING: memory at %lx maps "
397 "to invalid NUMA node %d\n", start,
402 if (max_domain < numa_domain)
403 max_domain = numa_domain;
405 if (! (size = numa_enforce_memory_limit(start, size))) {
413 * Initialize new node struct, or add to an existing one.
415 if (init_node_data[numa_domain].node_end_pfn) {
416 if ((start / PAGE_SIZE) <
417 init_node_data[numa_domain].node_start_pfn)
418 init_node_data[numa_domain].node_start_pfn =
420 if (((start / PAGE_SIZE) + (size / PAGE_SIZE)) >
421 init_node_data[numa_domain].node_end_pfn)
422 init_node_data[numa_domain].node_end_pfn =
423 (start / PAGE_SIZE) +
426 init_node_data[numa_domain].node_present_pages +=
429 node_set_online(numa_domain);
431 init_node_data[numa_domain].node_start_pfn =
433 init_node_data[numa_domain].node_end_pfn =
434 init_node_data[numa_domain].node_start_pfn +
436 init_node_data[numa_domain].node_present_pages =
440 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
441 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
448 for (i = 0; i <= max_domain; i++)
454 static void __init setup_nonnuma(void)
456 unsigned long top_of_ram = lmb_end_of_DRAM();
457 unsigned long total_ram = lmb_phys_mem_size();
460 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
461 top_of_ram, total_ram);
462 printk(KERN_INFO "Memory hole size: %ldMB\n",
463 (top_of_ram - total_ram) >> 20);
465 if (!numa_memory_lookup_table) {
466 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
467 numa_memory_lookup_table =
468 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
469 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
470 for (i = 0; i < entries ; i++)
471 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
474 map_cpu_to_node(boot_cpuid, 0);
478 init_node_data[0].node_start_pfn = 0;
479 init_node_data[0].node_end_pfn = lmb_end_of_DRAM() / PAGE_SIZE;
480 init_node_data[0].node_present_pages = total_ram / PAGE_SIZE;
482 for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
483 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
486 static void __init dump_numa_topology(void)
491 if (min_common_depth == -1 || !numa_enabled)
494 for_each_online_node(node) {
497 printk(KERN_INFO "Node %d Memory:", node);
501 for (i = 0; i < lmb_end_of_DRAM(); i += MEMORY_INCREMENT) {
502 if (numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] == node) {
521 * Allocate some memory, satisfying the lmb or bootmem allocator where
522 * required. nid is the preferred node and end is the physical address of
523 * the highest address in the node.
525 * Returns the physical address of the memory.
527 static unsigned long careful_allocation(int nid, unsigned long size,
528 unsigned long align, unsigned long end)
530 unsigned long ret = lmb_alloc_base(size, align, end);
532 /* retry over all memory */
534 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
537 panic("numa.c: cannot allocate %lu bytes on node %d",
541 * If the memory came from a previously allocated node, we must
542 * retry with the bootmem allocator.
544 if (pa_to_nid(ret) < nid) {
545 nid = pa_to_nid(ret);
546 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(nid),
550 panic("numa.c: cannot allocate %lu bytes on node %d",
553 ret = virt_to_abs(ret);
555 dbg("alloc_bootmem %lx %lx\n", ret, size);
561 void __init do_init_bootmem(void)
564 int addr_cells, size_cells;
565 struct device_node *memory = NULL;
566 static struct notifier_block ppc64_numa_nb = {
567 .notifier_call = cpu_numa_callback,
568 .priority = 1 /* Must run before sched domains notifier. */
572 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
573 max_pfn = max_low_pfn;
575 if (parse_numa_properties())
578 dump_numa_topology();
580 register_cpu_notifier(&ppc64_numa_nb);
582 for_each_online_node(nid) {
583 unsigned long start_paddr, end_paddr;
585 unsigned long bootmem_paddr;
586 unsigned long bootmap_pages;
588 start_paddr = init_node_data[nid].node_start_pfn * PAGE_SIZE;
589 end_paddr = init_node_data[nid].node_end_pfn * PAGE_SIZE;
591 /* Allocate the node structure node local if possible */
592 NODE_DATA(nid) = (struct pglist_data *)careful_allocation(nid,
593 sizeof(struct pglist_data),
594 SMP_CACHE_BYTES, end_paddr);
595 NODE_DATA(nid) = abs_to_virt(NODE_DATA(nid));
596 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
598 dbg("node %d\n", nid);
599 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
601 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
602 NODE_DATA(nid)->node_start_pfn =
603 init_node_data[nid].node_start_pfn;
604 NODE_DATA(nid)->node_spanned_pages =
605 end_paddr - start_paddr;
607 if (NODE_DATA(nid)->node_spanned_pages == 0)
610 dbg("start_paddr = %lx\n", start_paddr);
611 dbg("end_paddr = %lx\n", end_paddr);
613 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
615 bootmem_paddr = careful_allocation(nid,
616 bootmap_pages << PAGE_SHIFT,
617 PAGE_SIZE, end_paddr);
618 memset(abs_to_virt(bootmem_paddr), 0,
619 bootmap_pages << PAGE_SHIFT);
620 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
622 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
623 start_paddr >> PAGE_SHIFT,
624 end_paddr >> PAGE_SHIFT);
627 * We need to do another scan of all memory sections to
628 * associate memory with the correct node.
630 addr_cells = get_mem_addr_cells();
631 size_cells = get_mem_size_cells();
633 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
634 unsigned long mem_start, mem_size;
635 int numa_domain, ranges;
636 unsigned int *memcell_buf;
639 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
640 if (!memcell_buf || len <= 0)
643 ranges = memory->n_addrs; /* ranges in cell */
645 mem_start = read_n_cells(addr_cells, &memcell_buf);
646 mem_size = read_n_cells(size_cells, &memcell_buf);
648 numa_domain = of_node_numa_domain(memory);
649 if (numa_domain >= MAX_NUMNODES)
654 if (numa_domain != nid)
657 mem_size = numa_enforce_memory_limit(mem_start, mem_size);
659 dbg("free_bootmem %lx %lx\n", mem_start, mem_size);
660 free_bootmem_node(NODE_DATA(nid), mem_start, mem_size);
663 if (--ranges) /* process all ranges in cell */
668 * Mark reserved regions on this node
670 for (i = 0; i < lmb.reserved.cnt; i++) {
671 unsigned long physbase = lmb.reserved.region[i].base;
672 unsigned long size = lmb.reserved.region[i].size;
674 if (pa_to_nid(physbase) != nid &&
675 pa_to_nid(physbase+size-1) != nid)
678 if (physbase < end_paddr &&
679 (physbase+size) > start_paddr) {
681 if (physbase < start_paddr) {
682 size -= start_paddr - physbase;
683 physbase = start_paddr;
686 if (size > end_paddr - physbase)
687 size = end_paddr - physbase;
689 dbg("reserve_bootmem %lx %lx\n", physbase,
691 reserve_bootmem_node(NODE_DATA(nid), physbase,
696 * This loop may look famaliar, but we have to do it again
697 * after marking our reserved memory to mark memory present
700 addr_cells = get_mem_addr_cells();
701 size_cells = get_mem_size_cells();
703 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
704 unsigned long mem_start, mem_size;
705 int numa_domain, ranges;
706 unsigned int *memcell_buf;
709 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
710 if (!memcell_buf || len <= 0)
713 ranges = memory->n_addrs; /* ranges in cell */
715 mem_start = read_n_cells(addr_cells, &memcell_buf);
716 mem_size = read_n_cells(size_cells, &memcell_buf);
718 numa_domain = of_node_numa_domain(memory);
719 if (numa_domain >= MAX_NUMNODES)
724 if (numa_domain != nid)
727 mem_size = numa_enforce_memory_limit(mem_start, mem_size);
728 memory_present(numa_domain, mem_start >> PAGE_SHIFT,
729 (mem_start + mem_size) >> PAGE_SHIFT);
731 if (--ranges) /* process all ranges in cell */
738 void __init paging_init(void)
740 unsigned long zones_size[MAX_NR_ZONES];
741 unsigned long zholes_size[MAX_NR_ZONES];
744 memset(zones_size, 0, sizeof(zones_size));
745 memset(zholes_size, 0, sizeof(zholes_size));
747 for_each_online_node(nid) {
748 unsigned long start_pfn;
749 unsigned long end_pfn;
751 start_pfn = init_node_data[nid].node_start_pfn;
752 end_pfn = init_node_data[nid].node_end_pfn;
754 zones_size[ZONE_DMA] = end_pfn - start_pfn;
755 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
756 init_node_data[nid].node_present_pages;
758 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
759 zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
761 free_area_init_node(nid, NODE_DATA(nid), zones_size,
762 start_pfn, zholes_size);
766 static int __init early_numa(char *p)
771 if (strstr(p, "off"))
774 if (strstr(p, "debug"))
779 early_param("numa", early_numa);