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>
23 #include <asm/system.h>
26 static int numa_enabled = 1;
28 static int numa_debug;
29 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
32 #define ARRAY_INITIALISER -1
34 #define ARRAY_INITIALISER 0
37 int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
39 char *numa_memory_lookup_table;
40 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
41 int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
43 struct pglist_data *node_data[MAX_NUMNODES];
44 bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
45 static int min_common_depth;
48 * We need somewhere to store start/span for each node until we have
49 * allocated the real node_data structures.
52 unsigned long node_start_pfn;
53 unsigned long node_end_pfn;
54 unsigned long node_present_pages;
55 } init_node_data[MAX_NUMNODES] __initdata;
57 EXPORT_SYMBOL(node_data);
58 EXPORT_SYMBOL(numa_cpu_lookup_table);
59 EXPORT_SYMBOL(numa_memory_lookup_table);
60 EXPORT_SYMBOL(numa_cpumask_lookup_table);
61 EXPORT_SYMBOL(nr_cpus_in_node);
63 static inline void map_cpu_to_node(int cpu, int node)
65 numa_cpu_lookup_table[cpu] = node;
66 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node]))) {
67 cpu_set(cpu, numa_cpumask_lookup_table[node]);
68 nr_cpus_in_node[node]++;
72 #ifdef CONFIG_HOTPLUG_CPU
73 static void unmap_cpu_from_node(unsigned long cpu)
75 int node = numa_cpu_lookup_table[cpu];
77 dbg("removing cpu %lu from node %d\n", cpu, node);
79 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
80 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
81 nr_cpus_in_node[node]--;
83 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
87 #endif /* CONFIG_HOTPLUG_CPU */
89 static struct device_node * __devinit find_cpu_node(unsigned int cpu)
91 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
92 struct device_node *cpu_node = NULL;
93 unsigned int *interrupt_server, *reg;
96 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
97 /* Try interrupt server first */
98 interrupt_server = (unsigned int *)get_property(cpu_node,
99 "ibm,ppc-interrupt-server#s", &len);
101 len = len / sizeof(u32);
103 if (interrupt_server && (len > 0)) {
105 if (interrupt_server[len] == hw_cpuid)
109 reg = (unsigned int *)get_property(cpu_node,
111 if (reg && (len > 0) && (reg[0] == hw_cpuid))
119 /* must hold reference to node during call */
120 static int *of_get_associativity(struct device_node *dev)
122 return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
125 static int of_node_numa_domain(struct device_node *device)
130 if (min_common_depth == -1)
133 tmp = of_get_associativity(device);
134 if (tmp && (tmp[0] >= min_common_depth)) {
135 numa_domain = tmp[min_common_depth];
137 dbg("WARNING: no NUMA information for %s\n",
145 * In theory, the "ibm,associativity" property may contain multiple
146 * associativity lists because a resource may be multiply connected
147 * into the machine. This resource then has different associativity
148 * characteristics relative to its multiple connections. We ignore
149 * this for now. We also assume that all cpu and memory sets have
150 * their distances represented at a common level. This won't be
151 * true for heirarchical NUMA.
153 * In any case the ibm,associativity-reference-points should give
154 * the correct depth for a normal NUMA system.
156 * - Dave Hansen <haveblue@us.ibm.com>
158 static int __init find_min_common_depth(void)
161 unsigned int *ref_points;
162 struct device_node *rtas_root;
165 rtas_root = of_find_node_by_path("/rtas");
171 * this property is 2 32-bit integers, each representing a level of
172 * depth in the associativity nodes. The first is for an SMP
173 * configuration (should be all 0's) and the second is for a normal
174 * NUMA configuration.
176 ref_points = (unsigned int *)get_property(rtas_root,
177 "ibm,associativity-reference-points", &len);
179 if ((len >= 1) && ref_points) {
180 depth = ref_points[1];
182 dbg("WARNING: could not find NUMA "
183 "associativity reference point\n");
186 of_node_put(rtas_root);
191 static int __init get_mem_addr_cells(void)
193 struct device_node *memory = NULL;
196 memory = of_find_node_by_type(memory, "memory");
198 return 0; /* it won't matter */
200 rc = prom_n_addr_cells(memory);
204 static int __init get_mem_size_cells(void)
206 struct device_node *memory = NULL;
209 memory = of_find_node_by_type(memory, "memory");
211 return 0; /* it won't matter */
212 rc = prom_n_size_cells(memory);
216 static unsigned long read_n_cells(int n, unsigned int **buf)
218 unsigned long result = 0;
221 result = (result << 32) | **buf;
228 * Figure out to which domain a cpu belongs and stick it there.
229 * Return the id of the domain used.
231 static int numa_setup_cpu(unsigned long lcpu)
234 struct device_node *cpu = find_cpu_node(lcpu);
241 numa_domain = of_node_numa_domain(cpu);
243 if (numa_domain >= num_online_nodes()) {
245 * POWER4 LPAR uses 0xffff as invalid node,
246 * dont warn in this case.
248 if (numa_domain != 0xffff)
249 printk(KERN_ERR "WARNING: cpu %ld "
250 "maps to invalid NUMA node %d\n",
255 node_set_online(numa_domain);
257 map_cpu_to_node(lcpu, numa_domain);
264 static int cpu_numa_callback(struct notifier_block *nfb,
265 unsigned long action,
268 unsigned long lcpu = (unsigned long)hcpu;
269 int ret = NOTIFY_DONE;
273 if (min_common_depth == -1 || !numa_enabled)
274 map_cpu_to_node(lcpu, 0);
276 numa_setup_cpu(lcpu);
279 #ifdef CONFIG_HOTPLUG_CPU
281 case CPU_UP_CANCELED:
282 unmap_cpu_from_node(lcpu);
291 * Check and possibly modify a memory region to enforce the memory limit.
293 * Returns the size the region should have to enforce the memory limit.
294 * This will either be the original value of size, a truncated value,
295 * or zero. If the returned value of size is 0 the region should be
296 * discarded as it lies wholy above the memory limit.
298 static unsigned long __init numa_enforce_memory_limit(unsigned long start, unsigned long size)
301 * We use lmb_end_of_DRAM() in here instead of memory_limit because
302 * we've already adjusted it for the limit and it takes care of
303 * having memory holes below the limit.
309 if (start + size <= lmb_end_of_DRAM())
312 if (start >= lmb_end_of_DRAM())
315 return lmb_end_of_DRAM() - start;
318 static int __init parse_numa_properties(void)
320 struct device_node *cpu = NULL;
321 struct device_node *memory = NULL;
322 int addr_cells, size_cells;
324 long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
327 if (numa_enabled == 0) {
328 printk(KERN_WARNING "NUMA disabled by user\n");
332 numa_memory_lookup_table =
333 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
334 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
336 for (i = 0; i < entries ; i++)
337 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
339 min_common_depth = find_min_common_depth();
341 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
342 if (min_common_depth < 0)
343 return min_common_depth;
345 max_domain = numa_setup_cpu(boot_cpuid);
348 * Even though we connect cpus to numa domains later in SMP init,
349 * we need to know the maximum node id now. This is because each
350 * node id must have NODE_DATA etc backing it.
351 * As a result of hotplug we could still have cpus appear later on
352 * with larger node ids. In that case we force the cpu into node 0.
357 cpu = find_cpu_node(i);
360 numa_domain = of_node_numa_domain(cpu);
363 if (numa_domain < MAX_NUMNODES &&
364 max_domain < numa_domain)
365 max_domain = numa_domain;
369 addr_cells = get_mem_addr_cells();
370 size_cells = get_mem_size_cells();
372 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
377 unsigned int *memcell_buf;
380 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
381 if (!memcell_buf || len <= 0)
384 ranges = memory->n_addrs;
386 /* these are order-sensitive, and modify the buffer pointer */
387 start = read_n_cells(addr_cells, &memcell_buf);
388 size = read_n_cells(size_cells, &memcell_buf);
390 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
391 size = _ALIGN_UP(size, MEMORY_INCREMENT);
393 numa_domain = of_node_numa_domain(memory);
395 if (numa_domain >= MAX_NUMNODES) {
396 if (numa_domain != 0xffff)
397 printk(KERN_ERR "WARNING: memory at %lx maps "
398 "to invalid NUMA node %d\n", start,
403 if (max_domain < numa_domain)
404 max_domain = numa_domain;
406 if (! (size = numa_enforce_memory_limit(start, size))) {
414 * Initialize new node struct, or add to an existing one.
416 if (init_node_data[numa_domain].node_end_pfn) {
417 if ((start / PAGE_SIZE) <
418 init_node_data[numa_domain].node_start_pfn)
419 init_node_data[numa_domain].node_start_pfn =
421 if (((start / PAGE_SIZE) + (size / PAGE_SIZE)) >
422 init_node_data[numa_domain].node_end_pfn)
423 init_node_data[numa_domain].node_end_pfn =
424 (start / PAGE_SIZE) +
427 init_node_data[numa_domain].node_present_pages +=
430 node_set_online(numa_domain);
432 init_node_data[numa_domain].node_start_pfn =
434 init_node_data[numa_domain].node_end_pfn =
435 init_node_data[numa_domain].node_start_pfn +
437 init_node_data[numa_domain].node_present_pages =
441 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
442 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
449 for (i = 0; i <= max_domain; i++)
455 static void __init setup_nonnuma(void)
457 unsigned long top_of_ram = lmb_end_of_DRAM();
458 unsigned long total_ram = lmb_phys_mem_size();
461 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
462 top_of_ram, total_ram);
463 printk(KERN_INFO "Memory hole size: %ldMB\n",
464 (top_of_ram - total_ram) >> 20);
466 if (!numa_memory_lookup_table) {
467 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
468 numa_memory_lookup_table =
469 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
470 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
471 for (i = 0; i < entries ; i++)
472 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
475 map_cpu_to_node(boot_cpuid, 0);
479 init_node_data[0].node_start_pfn = 0;
480 init_node_data[0].node_end_pfn = lmb_end_of_DRAM() / PAGE_SIZE;
481 init_node_data[0].node_present_pages = total_ram / PAGE_SIZE;
483 for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
484 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
487 static void __init dump_numa_topology(void)
492 if (min_common_depth == -1 || !numa_enabled)
495 for_each_online_node(node) {
498 printk(KERN_INFO "Node %d Memory:", node);
502 for (i = 0; i < lmb_end_of_DRAM(); i += MEMORY_INCREMENT) {
503 if (numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] == node) {
522 * Allocate some memory, satisfying the lmb or bootmem allocator where
523 * required. nid is the preferred node and end is the physical address of
524 * the highest address in the node.
526 * Returns the physical address of the memory.
528 static unsigned long careful_allocation(int nid, unsigned long size,
529 unsigned long align, unsigned long end)
531 unsigned long ret = lmb_alloc_base(size, align, end);
533 /* retry over all memory */
535 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
538 panic("numa.c: cannot allocate %lu bytes on node %d",
542 * If the memory came from a previously allocated node, we must
543 * retry with the bootmem allocator.
545 if (pa_to_nid(ret) < nid) {
546 nid = pa_to_nid(ret);
547 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(nid),
551 panic("numa.c: cannot allocate %lu bytes on node %d",
554 ret = virt_to_abs(ret);
556 dbg("alloc_bootmem %lx %lx\n", ret, size);
562 void __init do_init_bootmem(void)
565 int addr_cells, size_cells;
566 struct device_node *memory = NULL;
567 static struct notifier_block ppc64_numa_nb = {
568 .notifier_call = cpu_numa_callback,
569 .priority = 1 /* Must run before sched domains notifier. */
573 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
574 max_pfn = max_low_pfn;
576 if (parse_numa_properties())
579 dump_numa_topology();
581 register_cpu_notifier(&ppc64_numa_nb);
583 for_each_online_node(nid) {
584 unsigned long start_paddr, end_paddr;
586 unsigned long bootmem_paddr;
587 unsigned long bootmap_pages;
589 start_paddr = init_node_data[nid].node_start_pfn * PAGE_SIZE;
590 end_paddr = init_node_data[nid].node_end_pfn * PAGE_SIZE;
592 /* Allocate the node structure node local if possible */
593 NODE_DATA(nid) = (struct pglist_data *)careful_allocation(nid,
594 sizeof(struct pglist_data),
595 SMP_CACHE_BYTES, end_paddr);
596 NODE_DATA(nid) = abs_to_virt(NODE_DATA(nid));
597 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
599 dbg("node %d\n", nid);
600 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
602 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
603 NODE_DATA(nid)->node_start_pfn =
604 init_node_data[nid].node_start_pfn;
605 NODE_DATA(nid)->node_spanned_pages =
606 end_paddr - start_paddr;
608 if (NODE_DATA(nid)->node_spanned_pages == 0)
611 dbg("start_paddr = %lx\n", start_paddr);
612 dbg("end_paddr = %lx\n", end_paddr);
614 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
616 bootmem_paddr = careful_allocation(nid,
617 bootmap_pages << PAGE_SHIFT,
618 PAGE_SIZE, end_paddr);
619 memset(abs_to_virt(bootmem_paddr), 0,
620 bootmap_pages << PAGE_SHIFT);
621 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
623 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
624 start_paddr >> PAGE_SHIFT,
625 end_paddr >> PAGE_SHIFT);
628 * We need to do another scan of all memory sections to
629 * associate memory with the correct node.
631 addr_cells = get_mem_addr_cells();
632 size_cells = get_mem_size_cells();
634 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
635 unsigned long mem_start, mem_size;
636 int numa_domain, ranges;
637 unsigned int *memcell_buf;
640 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
641 if (!memcell_buf || len <= 0)
644 ranges = memory->n_addrs; /* ranges in cell */
646 mem_start = read_n_cells(addr_cells, &memcell_buf);
647 mem_size = read_n_cells(size_cells, &memcell_buf);
649 numa_domain = of_node_numa_domain(memory);
650 if (numa_domain >= MAX_NUMNODES)
655 if (numa_domain != nid)
658 mem_size = numa_enforce_memory_limit(mem_start, mem_size);
660 dbg("free_bootmem %lx %lx\n", mem_start, mem_size);
661 free_bootmem_node(NODE_DATA(nid), mem_start, mem_size);
664 if (--ranges) /* process all ranges in cell */
669 * Mark reserved regions on this node
671 for (i = 0; i < lmb.reserved.cnt; i++) {
672 unsigned long physbase = lmb.reserved.region[i].base;
673 unsigned long size = lmb.reserved.region[i].size;
675 if (pa_to_nid(physbase) != nid &&
676 pa_to_nid(physbase+size-1) != nid)
679 if (physbase < end_paddr &&
680 (physbase+size) > start_paddr) {
682 if (physbase < start_paddr) {
683 size -= start_paddr - physbase;
684 physbase = start_paddr;
687 if (size > end_paddr - physbase)
688 size = end_paddr - physbase;
690 dbg("reserve_bootmem %lx %lx\n", physbase,
692 reserve_bootmem_node(NODE_DATA(nid), physbase,
697 * This loop may look famaliar, but we have to do it again
698 * after marking our reserved memory to mark memory present
701 addr_cells = get_mem_addr_cells();
702 size_cells = get_mem_size_cells();
704 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
705 unsigned long mem_start, mem_size;
706 int numa_domain, ranges;
707 unsigned int *memcell_buf;
710 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
711 if (!memcell_buf || len <= 0)
714 ranges = memory->n_addrs; /* ranges in cell */
716 mem_start = read_n_cells(addr_cells, &memcell_buf);
717 mem_size = read_n_cells(size_cells, &memcell_buf);
719 numa_domain = of_node_numa_domain(memory);
720 if (numa_domain >= MAX_NUMNODES)
725 if (numa_domain != nid)
728 mem_size = numa_enforce_memory_limit(mem_start, mem_size);
729 memory_present(numa_domain, mem_start >> PAGE_SHIFT,
730 (mem_start + mem_size) >> PAGE_SHIFT);
732 if (--ranges) /* process all ranges in cell */
739 void __init paging_init(void)
741 unsigned long zones_size[MAX_NR_ZONES];
742 unsigned long zholes_size[MAX_NR_ZONES];
745 memset(zones_size, 0, sizeof(zones_size));
746 memset(zholes_size, 0, sizeof(zholes_size));
748 for_each_online_node(nid) {
749 unsigned long start_pfn;
750 unsigned long end_pfn;
752 start_pfn = init_node_data[nid].node_start_pfn;
753 end_pfn = init_node_data[nid].node_end_pfn;
755 zones_size[ZONE_DMA] = end_pfn - start_pfn;
756 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
757 init_node_data[nid].node_present_pages;
759 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
760 zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
762 free_area_init_node(nid, NODE_DATA(nid), zones_size,
763 start_pfn, zholes_size);
767 static int __init early_numa(char *p)
772 if (strstr(p, "off"))
775 if (strstr(p, "debug"))
780 early_param("numa", early_numa);