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;
42 * We need somewhere to store start/end/node for each region until we have
43 * allocated the real node_data structures.
45 #define MAX_REGIONS (MAX_LMB_REGIONS*2)
47 unsigned long start_pfn;
48 unsigned long end_pfn;
50 } init_node_data[MAX_REGIONS] __initdata;
52 int __init early_pfn_to_nid(unsigned long pfn)
56 for (i = 0; init_node_data[i].end_pfn; i++) {
57 unsigned long start_pfn = init_node_data[i].start_pfn;
58 unsigned long end_pfn = init_node_data[i].end_pfn;
60 if ((start_pfn <= pfn) && (pfn < end_pfn))
61 return init_node_data[i].nid;
67 void __init add_region(unsigned int nid, unsigned long start_pfn,
72 dbg("add_region nid %d start_pfn 0x%lx pages 0x%lx\n",
73 nid, start_pfn, pages);
75 for (i = 0; init_node_data[i].end_pfn; i++) {
76 if (init_node_data[i].nid != nid)
78 if (init_node_data[i].end_pfn == start_pfn) {
79 init_node_data[i].end_pfn += pages;
82 if (init_node_data[i].start_pfn == (start_pfn + pages)) {
83 init_node_data[i].start_pfn -= pages;
89 * Leave last entry NULL so we dont iterate off the end (we use
90 * entry.end_pfn to terminate the walk).
92 if (i >= (MAX_REGIONS - 1)) {
93 printk(KERN_ERR "WARNING: too many memory regions in "
94 "numa code, truncating\n");
98 init_node_data[i].start_pfn = start_pfn;
99 init_node_data[i].end_pfn = start_pfn + pages;
100 init_node_data[i].nid = nid;
103 /* We assume init_node_data has no overlapping regions */
104 void __init get_region(unsigned int nid, unsigned long *start_pfn,
105 unsigned long *end_pfn, unsigned long *pages_present)
110 *end_pfn = *pages_present = 0;
112 for (i = 0; init_node_data[i].end_pfn; i++) {
113 if (init_node_data[i].nid != nid)
116 *pages_present += init_node_data[i].end_pfn -
117 init_node_data[i].start_pfn;
119 if (init_node_data[i].start_pfn < *start_pfn)
120 *start_pfn = init_node_data[i].start_pfn;
122 if (init_node_data[i].end_pfn > *end_pfn)
123 *end_pfn = init_node_data[i].end_pfn;
126 /* We didnt find a matching region, return start/end as 0 */
127 if (*start_pfn == -1UL)
131 static inline void map_cpu_to_node(int cpu, int node)
133 numa_cpu_lookup_table[cpu] = node;
135 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
136 cpu_set(cpu, numa_cpumask_lookup_table[node]);
139 #ifdef CONFIG_HOTPLUG_CPU
140 static void unmap_cpu_from_node(unsigned long cpu)
142 int node = numa_cpu_lookup_table[cpu];
144 dbg("removing cpu %lu from node %d\n", cpu, node);
146 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
147 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
149 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
153 #endif /* CONFIG_HOTPLUG_CPU */
155 static struct device_node *find_cpu_node(unsigned int cpu)
157 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
158 struct device_node *cpu_node = NULL;
159 unsigned int *interrupt_server, *reg;
162 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
163 /* Try interrupt server first */
164 interrupt_server = (unsigned int *)get_property(cpu_node,
165 "ibm,ppc-interrupt-server#s", &len);
167 len = len / sizeof(u32);
169 if (interrupt_server && (len > 0)) {
171 if (interrupt_server[len] == hw_cpuid)
175 reg = (unsigned int *)get_property(cpu_node,
177 if (reg && (len > 0) && (reg[0] == hw_cpuid))
185 /* must hold reference to node during call */
186 static int *of_get_associativity(struct device_node *dev)
188 return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
191 static int of_node_numa_domain(struct device_node *device)
196 if (min_common_depth == -1)
199 tmp = of_get_associativity(device);
200 if (tmp && (tmp[0] >= min_common_depth)) {
201 numa_domain = tmp[min_common_depth];
203 dbg("WARNING: no NUMA information for %s\n",
211 * In theory, the "ibm,associativity" property may contain multiple
212 * associativity lists because a resource may be multiply connected
213 * into the machine. This resource then has different associativity
214 * characteristics relative to its multiple connections. We ignore
215 * this for now. We also assume that all cpu and memory sets have
216 * their distances represented at a common level. This won't be
217 * true for heirarchical NUMA.
219 * In any case the ibm,associativity-reference-points should give
220 * the correct depth for a normal NUMA system.
222 * - Dave Hansen <haveblue@us.ibm.com>
224 static int __init find_min_common_depth(void)
227 unsigned int *ref_points;
228 struct device_node *rtas_root;
231 rtas_root = of_find_node_by_path("/rtas");
237 * this property is 2 32-bit integers, each representing a level of
238 * depth in the associativity nodes. The first is for an SMP
239 * configuration (should be all 0's) and the second is for a normal
240 * NUMA configuration.
242 ref_points = (unsigned int *)get_property(rtas_root,
243 "ibm,associativity-reference-points", &len);
245 if ((len >= 1) && ref_points) {
246 depth = ref_points[1];
248 dbg("WARNING: could not find NUMA "
249 "associativity reference point\n");
252 of_node_put(rtas_root);
257 static int __init get_mem_addr_cells(void)
259 struct device_node *memory = NULL;
262 memory = of_find_node_by_type(memory, "memory");
264 return 0; /* it won't matter */
266 rc = prom_n_addr_cells(memory);
270 static int __init get_mem_size_cells(void)
272 struct device_node *memory = NULL;
275 memory = of_find_node_by_type(memory, "memory");
277 return 0; /* it won't matter */
278 rc = prom_n_size_cells(memory);
282 static unsigned long __init read_n_cells(int n, unsigned int **buf)
284 unsigned long result = 0;
287 result = (result << 32) | **buf;
294 * Figure out to which domain a cpu belongs and stick it there.
295 * Return the id of the domain used.
297 static int numa_setup_cpu(unsigned long lcpu)
300 struct device_node *cpu = find_cpu_node(lcpu);
307 numa_domain = of_node_numa_domain(cpu);
309 if (numa_domain >= num_online_nodes()) {
311 * POWER4 LPAR uses 0xffff as invalid node,
312 * dont warn in this case.
314 if (numa_domain != 0xffff)
315 printk(KERN_ERR "WARNING: cpu %ld "
316 "maps to invalid NUMA node %d\n",
321 node_set_online(numa_domain);
323 map_cpu_to_node(lcpu, numa_domain);
330 static int cpu_numa_callback(struct notifier_block *nfb,
331 unsigned long action,
334 unsigned long lcpu = (unsigned long)hcpu;
335 int ret = NOTIFY_DONE;
339 if (min_common_depth == -1 || !numa_enabled)
340 map_cpu_to_node(lcpu, 0);
342 numa_setup_cpu(lcpu);
345 #ifdef CONFIG_HOTPLUG_CPU
347 case CPU_UP_CANCELED:
348 unmap_cpu_from_node(lcpu);
357 * Check and possibly modify a memory region to enforce the memory limit.
359 * Returns the size the region should have to enforce the memory limit.
360 * This will either be the original value of size, a truncated value,
361 * or zero. If the returned value of size is 0 the region should be
362 * discarded as it lies wholy above the memory limit.
364 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
368 * We use lmb_end_of_DRAM() in here instead of memory_limit because
369 * we've already adjusted it for the limit and it takes care of
370 * having memory holes below the limit.
376 if (start + size <= lmb_end_of_DRAM())
379 if (start >= lmb_end_of_DRAM())
382 return lmb_end_of_DRAM() - start;
385 static int __init parse_numa_properties(void)
387 struct device_node *cpu = NULL;
388 struct device_node *memory = NULL;
389 int addr_cells, size_cells;
393 if (numa_enabled == 0) {
394 printk(KERN_WARNING "NUMA disabled by user\n");
398 min_common_depth = find_min_common_depth();
400 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
401 if (min_common_depth < 0)
402 return min_common_depth;
404 max_domain = numa_setup_cpu(boot_cpuid);
407 * Even though we connect cpus to numa domains later in SMP init,
408 * we need to know the maximum node id now. This is because each
409 * node id must have NODE_DATA etc backing it.
410 * As a result of hotplug we could still have cpus appear later on
411 * with larger node ids. In that case we force the cpu into node 0.
416 cpu = find_cpu_node(i);
419 numa_domain = of_node_numa_domain(cpu);
422 if (numa_domain < MAX_NUMNODES &&
423 max_domain < numa_domain)
424 max_domain = numa_domain;
428 addr_cells = get_mem_addr_cells();
429 size_cells = get_mem_size_cells();
431 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
436 unsigned int *memcell_buf;
439 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
440 if (!memcell_buf || len <= 0)
443 ranges = memory->n_addrs;
445 /* these are order-sensitive, and modify the buffer pointer */
446 start = read_n_cells(addr_cells, &memcell_buf);
447 size = read_n_cells(size_cells, &memcell_buf);
449 numa_domain = of_node_numa_domain(memory);
451 if (numa_domain >= MAX_NUMNODES) {
452 if (numa_domain != 0xffff)
453 printk(KERN_ERR "WARNING: memory at %lx maps "
454 "to invalid NUMA node %d\n", start,
459 if (max_domain < numa_domain)
460 max_domain = numa_domain;
462 if (!(size = numa_enforce_memory_limit(start, size))) {
469 add_region(numa_domain, start >> PAGE_SHIFT,
476 for (i = 0; i <= max_domain; i++)
482 static void __init setup_nonnuma(void)
484 unsigned long top_of_ram = lmb_end_of_DRAM();
485 unsigned long total_ram = lmb_phys_mem_size();
488 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
489 top_of_ram, total_ram);
490 printk(KERN_INFO "Memory hole size: %ldMB\n",
491 (top_of_ram - total_ram) >> 20);
493 map_cpu_to_node(boot_cpuid, 0);
494 for (i = 0; i < lmb.memory.cnt; ++i)
495 add_region(0, lmb.memory.region[i].base >> PAGE_SHIFT,
496 lmb_size_pages(&lmb.memory, i));
500 static void __init dump_numa_topology(void)
505 if (min_common_depth == -1 || !numa_enabled)
508 for_each_online_node(node) {
511 printk(KERN_INFO "Node %d Memory:", node);
515 for (i = 0; i < lmb_end_of_DRAM();
516 i += (1 << SECTION_SIZE_BITS)) {
517 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
536 * Allocate some memory, satisfying the lmb or bootmem allocator where
537 * required. nid is the preferred node and end is the physical address of
538 * the highest address in the node.
540 * Returns the physical address of the memory.
542 static void __init *careful_allocation(int nid, unsigned long size,
544 unsigned long end_pfn)
547 unsigned long ret = lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
549 /* retry over all memory */
551 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
554 panic("numa.c: cannot allocate %lu bytes on node %d",
558 * If the memory came from a previously allocated node, we must
559 * retry with the bootmem allocator.
561 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
563 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
567 panic("numa.c: cannot allocate %lu bytes on node %d",
572 dbg("alloc_bootmem %lx %lx\n", ret, size);
578 void __init do_init_bootmem(void)
582 static struct notifier_block ppc64_numa_nb = {
583 .notifier_call = cpu_numa_callback,
584 .priority = 1 /* Must run before sched domains notifier. */
588 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
589 max_pfn = max_low_pfn;
591 if (parse_numa_properties())
594 dump_numa_topology();
596 register_cpu_notifier(&ppc64_numa_nb);
598 for_each_online_node(nid) {
599 unsigned long start_pfn, end_pfn, pages_present;
600 unsigned long bootmem_paddr;
601 unsigned long bootmap_pages;
603 get_region(nid, &start_pfn, &end_pfn, &pages_present);
605 /* Allocate the node structure node local if possible */
606 NODE_DATA(nid) = careful_allocation(nid,
607 sizeof(struct pglist_data),
608 SMP_CACHE_BYTES, end_pfn);
609 NODE_DATA(nid) = __va(NODE_DATA(nid));
610 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
612 dbg("node %d\n", nid);
613 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
615 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
616 NODE_DATA(nid)->node_start_pfn = start_pfn;
617 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
619 if (NODE_DATA(nid)->node_spanned_pages == 0)
622 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
623 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
625 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
626 bootmem_paddr = (unsigned long)careful_allocation(nid,
627 bootmap_pages << PAGE_SHIFT,
629 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
631 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
633 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
636 /* Add free regions on this node */
637 for (i = 0; init_node_data[i].end_pfn; i++) {
638 unsigned long start, end;
640 if (init_node_data[i].nid != nid)
643 start = init_node_data[i].start_pfn << PAGE_SHIFT;
644 end = init_node_data[i].end_pfn << PAGE_SHIFT;
646 dbg("free_bootmem %lx %lx\n", start, end - start);
647 free_bootmem_node(NODE_DATA(nid), start, end - start);
650 /* Mark reserved regions on this node */
651 for (i = 0; i < lmb.reserved.cnt; i++) {
652 unsigned long physbase = lmb.reserved.region[i].base;
653 unsigned long size = lmb.reserved.region[i].size;
654 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
655 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
657 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
658 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
661 if (physbase < end_paddr &&
662 (physbase+size) > start_paddr) {
664 if (physbase < start_paddr) {
665 size -= start_paddr - physbase;
666 physbase = start_paddr;
669 if (size > end_paddr - physbase)
670 size = end_paddr - physbase;
672 dbg("reserve_bootmem %lx %lx\n", physbase,
674 reserve_bootmem_node(NODE_DATA(nid), physbase,
679 /* Add regions into sparsemem */
680 for (i = 0; init_node_data[i].end_pfn; i++) {
681 unsigned long start, end;
683 if (init_node_data[i].nid != nid)
686 start = init_node_data[i].start_pfn;
687 end = init_node_data[i].end_pfn;
689 memory_present(nid, start, end);
694 void __init paging_init(void)
696 unsigned long zones_size[MAX_NR_ZONES];
697 unsigned long zholes_size[MAX_NR_ZONES];
700 memset(zones_size, 0, sizeof(zones_size));
701 memset(zholes_size, 0, sizeof(zholes_size));
703 for_each_online_node(nid) {
704 unsigned long start_pfn, end_pfn, pages_present;
706 get_region(nid, &start_pfn, &end_pfn, &pages_present);
708 zones_size[ZONE_DMA] = end_pfn - start_pfn;
709 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] - pages_present;
711 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
712 zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
714 free_area_init_node(nid, NODE_DATA(nid), zones_size, start_pfn,
719 static int __init early_numa(char *p)
724 if (strstr(p, "off"))
727 if (strstr(p, "debug"))
732 early_param("numa", early_numa);