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;
43 * We need somewhere to store start/end/node for each region until we have
44 * allocated the real node_data structures.
46 #define MAX_REGIONS (MAX_LMB_REGIONS*2)
48 unsigned long start_pfn;
49 unsigned long end_pfn;
51 } init_node_data[MAX_REGIONS] __initdata;
53 int __init early_pfn_to_nid(unsigned long pfn)
57 for (i = 0; init_node_data[i].end_pfn; i++) {
58 unsigned long start_pfn = init_node_data[i].start_pfn;
59 unsigned long end_pfn = init_node_data[i].end_pfn;
61 if ((start_pfn <= pfn) && (pfn < end_pfn))
62 return init_node_data[i].nid;
68 void __init add_region(unsigned int nid, unsigned long start_pfn,
73 dbg("add_region nid %d start_pfn 0x%lx pages 0x%lx\n",
74 nid, start_pfn, pages);
76 for (i = 0; init_node_data[i].end_pfn; i++) {
77 if (init_node_data[i].nid != nid)
79 if (init_node_data[i].end_pfn == start_pfn) {
80 init_node_data[i].end_pfn += pages;
83 if (init_node_data[i].start_pfn == (start_pfn + pages)) {
84 init_node_data[i].start_pfn -= pages;
90 * Leave last entry NULL so we dont iterate off the end (we use
91 * entry.end_pfn to terminate the walk).
93 if (i >= (MAX_REGIONS - 1)) {
94 printk(KERN_ERR "WARNING: too many memory regions in "
95 "numa code, truncating\n");
99 init_node_data[i].start_pfn = start_pfn;
100 init_node_data[i].end_pfn = start_pfn + pages;
101 init_node_data[i].nid = nid;
104 /* We assume init_node_data has no overlapping regions */
105 void __init get_region(unsigned int nid, unsigned long *start_pfn,
106 unsigned long *end_pfn, unsigned long *pages_present)
111 *end_pfn = *pages_present = 0;
113 for (i = 0; init_node_data[i].end_pfn; i++) {
114 if (init_node_data[i].nid != nid)
117 *pages_present += init_node_data[i].end_pfn -
118 init_node_data[i].start_pfn;
120 if (init_node_data[i].start_pfn < *start_pfn)
121 *start_pfn = init_node_data[i].start_pfn;
123 if (init_node_data[i].end_pfn > *end_pfn)
124 *end_pfn = init_node_data[i].end_pfn;
127 /* We didnt find a matching region, return start/end as 0 */
128 if (*start_pfn == -1UL)
132 static inline void map_cpu_to_node(int cpu, int node)
134 numa_cpu_lookup_table[cpu] = node;
136 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
137 cpu_set(cpu, numa_cpumask_lookup_table[node]);
140 #ifdef CONFIG_HOTPLUG_CPU
141 static void unmap_cpu_from_node(unsigned long cpu)
143 int node = numa_cpu_lookup_table[cpu];
145 dbg("removing cpu %lu from node %d\n", cpu, node);
147 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
148 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
150 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
154 #endif /* CONFIG_HOTPLUG_CPU */
156 static struct device_node *find_cpu_node(unsigned int cpu)
158 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
159 struct device_node *cpu_node = NULL;
160 unsigned int *interrupt_server, *reg;
163 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
164 /* Try interrupt server first */
165 interrupt_server = (unsigned int *)get_property(cpu_node,
166 "ibm,ppc-interrupt-server#s", &len);
168 len = len / sizeof(u32);
170 if (interrupt_server && (len > 0)) {
172 if (interrupt_server[len] == hw_cpuid)
176 reg = (unsigned int *)get_property(cpu_node,
178 if (reg && (len > 0) && (reg[0] == hw_cpuid))
186 /* must hold reference to node during call */
187 static int *of_get_associativity(struct device_node *dev)
189 return (unsigned int *)get_property(dev, "ibm,associativity", NULL);
192 static int of_node_numa_domain(struct device_node *device)
197 if (min_common_depth == -1)
200 tmp = of_get_associativity(device);
201 if (tmp && (tmp[0] >= min_common_depth)) {
202 numa_domain = tmp[min_common_depth];
204 dbg("WARNING: no NUMA information for %s\n",
212 * In theory, the "ibm,associativity" property may contain multiple
213 * associativity lists because a resource may be multiply connected
214 * into the machine. This resource then has different associativity
215 * characteristics relative to its multiple connections. We ignore
216 * this for now. We also assume that all cpu and memory sets have
217 * their distances represented at a common level. This won't be
218 * true for heirarchical NUMA.
220 * In any case the ibm,associativity-reference-points should give
221 * the correct depth for a normal NUMA system.
223 * - Dave Hansen <haveblue@us.ibm.com>
225 static int __init find_min_common_depth(void)
228 unsigned int *ref_points;
229 struct device_node *rtas_root;
232 rtas_root = of_find_node_by_path("/rtas");
238 * this property is 2 32-bit integers, each representing a level of
239 * depth in the associativity nodes. The first is for an SMP
240 * configuration (should be all 0's) and the second is for a normal
241 * NUMA configuration.
243 ref_points = (unsigned int *)get_property(rtas_root,
244 "ibm,associativity-reference-points", &len);
246 if ((len >= 1) && ref_points) {
247 depth = ref_points[1];
249 dbg("WARNING: could not find NUMA "
250 "associativity reference point\n");
253 of_node_put(rtas_root);
258 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
260 struct device_node *memory = NULL;
262 memory = of_find_node_by_type(memory, "memory");
264 panic("numa.c: No memory nodes found!");
266 *n_addr_cells = prom_n_addr_cells(memory);
267 *n_size_cells = prom_n_size_cells(memory);
271 static unsigned long __devinit read_n_cells(int n, unsigned int **buf)
273 unsigned long result = 0;
276 result = (result << 32) | **buf;
283 * Figure out to which domain a cpu belongs and stick it there.
284 * Return the id of the domain used.
286 static int numa_setup_cpu(unsigned long lcpu)
289 struct device_node *cpu = find_cpu_node(lcpu);
296 numa_domain = of_node_numa_domain(cpu);
298 if (numa_domain >= num_online_nodes()) {
300 * POWER4 LPAR uses 0xffff as invalid node,
301 * dont warn in this case.
303 if (numa_domain != 0xffff)
304 printk(KERN_ERR "WARNING: cpu %ld "
305 "maps to invalid NUMA node %d\n",
310 node_set_online(numa_domain);
312 map_cpu_to_node(lcpu, numa_domain);
319 static int cpu_numa_callback(struct notifier_block *nfb,
320 unsigned long action,
323 unsigned long lcpu = (unsigned long)hcpu;
324 int ret = NOTIFY_DONE;
328 if (min_common_depth == -1 || !numa_enabled)
329 map_cpu_to_node(lcpu, 0);
331 numa_setup_cpu(lcpu);
334 #ifdef CONFIG_HOTPLUG_CPU
336 case CPU_UP_CANCELED:
337 unmap_cpu_from_node(lcpu);
346 * Check and possibly modify a memory region to enforce the memory limit.
348 * Returns the size the region should have to enforce the memory limit.
349 * This will either be the original value of size, a truncated value,
350 * or zero. If the returned value of size is 0 the region should be
351 * discarded as it lies wholy above the memory limit.
353 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
357 * We use lmb_end_of_DRAM() in here instead of memory_limit because
358 * we've already adjusted it for the limit and it takes care of
359 * having memory holes below the limit.
365 if (start + size <= lmb_end_of_DRAM())
368 if (start >= lmb_end_of_DRAM())
371 return lmb_end_of_DRAM() - start;
374 static int __init parse_numa_properties(void)
376 struct device_node *cpu = NULL;
377 struct device_node *memory = NULL;
381 if (numa_enabled == 0) {
382 printk(KERN_WARNING "NUMA disabled by user\n");
386 min_common_depth = find_min_common_depth();
388 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
389 if (min_common_depth < 0)
390 return min_common_depth;
392 max_domain = numa_setup_cpu(boot_cpuid);
395 * Even though we connect cpus to numa domains later in SMP init,
396 * we need to know the maximum node id now. This is because each
397 * node id must have NODE_DATA etc backing it.
398 * As a result of hotplug we could still have cpus appear later on
399 * with larger node ids. In that case we force the cpu into node 0.
404 cpu = find_cpu_node(i);
407 numa_domain = of_node_numa_domain(cpu);
410 if (numa_domain < MAX_NUMNODES &&
411 max_domain < numa_domain)
412 max_domain = numa_domain;
416 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
418 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
423 unsigned int *memcell_buf;
426 memcell_buf = (unsigned int *)get_property(memory,
427 "linux,usable-memory", &len);
428 if (!memcell_buf || len <= 0)
430 (unsigned int *)get_property(memory, "reg",
432 if (!memcell_buf || len <= 0)
435 ranges = memory->n_addrs;
437 /* these are order-sensitive, and modify the buffer pointer */
438 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
439 size = read_n_cells(n_mem_size_cells, &memcell_buf);
441 numa_domain = of_node_numa_domain(memory);
443 if (numa_domain >= MAX_NUMNODES) {
444 if (numa_domain != 0xffff)
445 printk(KERN_ERR "WARNING: memory at %lx maps "
446 "to invalid NUMA node %d\n", start,
451 if (max_domain < numa_domain)
452 max_domain = numa_domain;
454 if (!(size = numa_enforce_memory_limit(start, size))) {
461 add_region(numa_domain, start >> PAGE_SHIFT,
468 for (i = 0; i <= max_domain; i++)
474 static void __init setup_nonnuma(void)
476 unsigned long top_of_ram = lmb_end_of_DRAM();
477 unsigned long total_ram = lmb_phys_mem_size();
480 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
481 top_of_ram, total_ram);
482 printk(KERN_INFO "Memory hole size: %ldMB\n",
483 (top_of_ram - total_ram) >> 20);
485 map_cpu_to_node(boot_cpuid, 0);
486 for (i = 0; i < lmb.memory.cnt; ++i)
487 add_region(0, lmb.memory.region[i].base >> PAGE_SHIFT,
488 lmb_size_pages(&lmb.memory, i));
492 static void __init dump_numa_topology(void)
497 if (min_common_depth == -1 || !numa_enabled)
500 for_each_online_node(node) {
503 printk(KERN_INFO "Node %d Memory:", node);
507 for (i = 0; i < lmb_end_of_DRAM();
508 i += (1 << SECTION_SIZE_BITS)) {
509 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
528 * Allocate some memory, satisfying the lmb or bootmem allocator where
529 * required. nid is the preferred node and end is the physical address of
530 * the highest address in the node.
532 * Returns the physical address of the memory.
534 static void __init *careful_allocation(int nid, unsigned long size,
536 unsigned long end_pfn)
539 unsigned long ret = lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
541 /* retry over all memory */
543 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
546 panic("numa.c: cannot allocate %lu bytes on node %d",
550 * If the memory came from a previously allocated node, we must
551 * retry with the bootmem allocator.
553 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
555 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
559 panic("numa.c: cannot allocate %lu bytes on node %d",
564 dbg("alloc_bootmem %lx %lx\n", ret, size);
570 void __init do_init_bootmem(void)
574 static struct notifier_block ppc64_numa_nb = {
575 .notifier_call = cpu_numa_callback,
576 .priority = 1 /* Must run before sched domains notifier. */
580 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
581 max_pfn = max_low_pfn;
583 if (parse_numa_properties())
586 dump_numa_topology();
588 register_cpu_notifier(&ppc64_numa_nb);
590 for_each_online_node(nid) {
591 unsigned long start_pfn, end_pfn, pages_present;
592 unsigned long bootmem_paddr;
593 unsigned long bootmap_pages;
595 get_region(nid, &start_pfn, &end_pfn, &pages_present);
597 /* Allocate the node structure node local if possible */
598 NODE_DATA(nid) = careful_allocation(nid,
599 sizeof(struct pglist_data),
600 SMP_CACHE_BYTES, end_pfn);
601 NODE_DATA(nid) = __va(NODE_DATA(nid));
602 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
604 dbg("node %d\n", nid);
605 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
607 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
608 NODE_DATA(nid)->node_start_pfn = start_pfn;
609 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
611 if (NODE_DATA(nid)->node_spanned_pages == 0)
614 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
615 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
617 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
618 bootmem_paddr = (unsigned long)careful_allocation(nid,
619 bootmap_pages << PAGE_SHIFT,
621 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
623 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
625 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
628 /* Add free regions on this node */
629 for (i = 0; init_node_data[i].end_pfn; i++) {
630 unsigned long start, end;
632 if (init_node_data[i].nid != nid)
635 start = init_node_data[i].start_pfn << PAGE_SHIFT;
636 end = init_node_data[i].end_pfn << PAGE_SHIFT;
638 dbg("free_bootmem %lx %lx\n", start, end - start);
639 free_bootmem_node(NODE_DATA(nid), start, end - start);
642 /* Mark reserved regions on this node */
643 for (i = 0; i < lmb.reserved.cnt; i++) {
644 unsigned long physbase = lmb.reserved.region[i].base;
645 unsigned long size = lmb.reserved.region[i].size;
646 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
647 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
649 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
650 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
653 if (physbase < end_paddr &&
654 (physbase+size) > start_paddr) {
656 if (physbase < start_paddr) {
657 size -= start_paddr - physbase;
658 physbase = start_paddr;
661 if (size > end_paddr - physbase)
662 size = end_paddr - physbase;
664 dbg("reserve_bootmem %lx %lx\n", physbase,
666 reserve_bootmem_node(NODE_DATA(nid), physbase,
671 /* Add regions into sparsemem */
672 for (i = 0; init_node_data[i].end_pfn; i++) {
673 unsigned long start, end;
675 if (init_node_data[i].nid != nid)
678 start = init_node_data[i].start_pfn;
679 end = init_node_data[i].end_pfn;
681 memory_present(nid, start, end);
686 void __init paging_init(void)
688 unsigned long zones_size[MAX_NR_ZONES];
689 unsigned long zholes_size[MAX_NR_ZONES];
692 memset(zones_size, 0, sizeof(zones_size));
693 memset(zholes_size, 0, sizeof(zholes_size));
695 for_each_online_node(nid) {
696 unsigned long start_pfn, end_pfn, pages_present;
698 get_region(nid, &start_pfn, &end_pfn, &pages_present);
700 zones_size[ZONE_DMA] = end_pfn - start_pfn;
701 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] - pages_present;
703 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
704 zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
706 free_area_init_node(nid, NODE_DATA(nid), zones_size, start_pfn,
711 static int __init early_numa(char *p)
716 if (strstr(p, "off"))
719 if (strstr(p, "debug"))
724 early_param("numa", early_numa);
726 #ifdef CONFIG_MEMORY_HOTPLUG
728 * Find the node associated with a hot added memory section. Section
729 * corresponds to a SPARSEMEM section, not an LMB. It is assumed that
730 * sections are fully contained within a single LMB.
732 int hot_add_scn_to_nid(unsigned long scn_addr)
734 struct device_node *memory = NULL;
736 if (!numa_enabled || (min_common_depth < 0))
739 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
740 unsigned long start, size;
741 int numa_domain, ranges;
742 unsigned int *memcell_buf;
745 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
746 if (!memcell_buf || len <= 0)
749 ranges = memory->n_addrs; /* ranges in cell */
751 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
752 size = read_n_cells(n_mem_size_cells, &memcell_buf);
753 numa_domain = of_node_numa_domain(memory);
755 /* Domains not present at boot default to 0 */
756 if (!node_online(numa_domain))
757 numa_domain = any_online_node(NODE_MASK_ALL);
759 if ((scn_addr >= start) && (scn_addr < (start + size))) {
764 if (--ranges) /* process all ranges in cell */
768 BUG(); /* section address should be found above */
771 #endif /* CONFIG_MEMORY_HOTPLUG */