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 <linux/lmb.h>
21 #include <asm/sparsemem.h>
23 #include <asm/system.h>
26 static int numa_enabled = 1;
28 static char *cmdline __initdata;
30 static int numa_debug;
31 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
33 int numa_cpu_lookup_table[NR_CPUS];
34 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
35 struct pglist_data *node_data[MAX_NUMNODES];
37 EXPORT_SYMBOL(numa_cpu_lookup_table);
38 EXPORT_SYMBOL(numa_cpumask_lookup_table);
39 EXPORT_SYMBOL(node_data);
41 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
42 static int min_common_depth;
43 static int n_mem_addr_cells, n_mem_size_cells;
45 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
48 unsigned long long mem;
50 static unsigned int fake_nid;
51 static unsigned long long curr_boundary;
54 * Modify node id, iff we started creating NUMA nodes
55 * We want to continue from where we left of the last time
60 * In case there are no more arguments to parse, the
61 * node_id should be the same as the last fake node id
62 * (we've handled this above).
67 mem = memparse(p, &p);
71 if (mem < curr_boundary)
76 if ((end_pfn << PAGE_SHIFT) > mem) {
78 * Skip commas and spaces
80 while (*p == ',' || *p == ' ' || *p == '\t')
86 dbg("created new fake_node with id %d\n", fake_nid);
92 static void __cpuinit map_cpu_to_node(int cpu, int node)
94 numa_cpu_lookup_table[cpu] = node;
96 dbg("adding cpu %d to node %d\n", cpu, node);
98 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
99 cpu_set(cpu, numa_cpumask_lookup_table[node]);
102 #ifdef CONFIG_HOTPLUG_CPU
103 static void unmap_cpu_from_node(unsigned long cpu)
105 int node = numa_cpu_lookup_table[cpu];
107 dbg("removing cpu %lu from node %d\n", cpu, node);
109 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
110 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
112 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
116 #endif /* CONFIG_HOTPLUG_CPU */
118 static struct device_node * __cpuinit find_cpu_node(unsigned int cpu)
120 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
121 struct device_node *cpu_node = NULL;
122 const unsigned int *interrupt_server, *reg;
125 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
126 /* Try interrupt server first */
127 interrupt_server = of_get_property(cpu_node,
128 "ibm,ppc-interrupt-server#s", &len);
130 len = len / sizeof(u32);
132 if (interrupt_server && (len > 0)) {
134 if (interrupt_server[len] == hw_cpuid)
138 reg = of_get_property(cpu_node, "reg", &len);
139 if (reg && (len > 0) && (reg[0] == hw_cpuid))
147 /* must hold reference to node during call */
148 static const int *of_get_associativity(struct device_node *dev)
150 return of_get_property(dev, "ibm,associativity", NULL);
153 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
156 static int of_node_to_nid_single(struct device_node *device)
159 const unsigned int *tmp;
161 if (min_common_depth == -1)
164 tmp = of_get_associativity(device);
168 if (tmp[0] >= min_common_depth)
169 nid = tmp[min_common_depth];
171 /* POWER4 LPAR uses 0xffff as invalid node */
172 if (nid == 0xffff || nid >= MAX_NUMNODES)
178 /* Walk the device tree upwards, looking for an associativity id */
179 int of_node_to_nid(struct device_node *device)
181 struct device_node *tmp;
186 nid = of_node_to_nid_single(device);
191 device = of_get_parent(tmp);
198 EXPORT_SYMBOL_GPL(of_node_to_nid);
201 * In theory, the "ibm,associativity" property may contain multiple
202 * associativity lists because a resource may be multiply connected
203 * into the machine. This resource then has different associativity
204 * characteristics relative to its multiple connections. We ignore
205 * this for now. We also assume that all cpu and memory sets have
206 * their distances represented at a common level. This won't be
207 * true for hierarchical NUMA.
209 * In any case the ibm,associativity-reference-points should give
210 * the correct depth for a normal NUMA system.
212 * - Dave Hansen <haveblue@us.ibm.com>
214 static int __init find_min_common_depth(void)
217 const unsigned int *ref_points;
218 struct device_node *rtas_root;
221 rtas_root = of_find_node_by_path("/rtas");
227 * this property is 2 32-bit integers, each representing a level of
228 * depth in the associativity nodes. The first is for an SMP
229 * configuration (should be all 0's) and the second is for a normal
230 * NUMA configuration.
232 ref_points = of_get_property(rtas_root,
233 "ibm,associativity-reference-points", &len);
235 if ((len >= 1) && ref_points) {
236 depth = ref_points[1];
238 dbg("NUMA: ibm,associativity-reference-points not found.\n");
241 of_node_put(rtas_root);
246 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
248 struct device_node *memory = NULL;
250 memory = of_find_node_by_type(memory, "memory");
252 panic("numa.c: No memory nodes found!");
254 *n_addr_cells = of_n_addr_cells(memory);
255 *n_size_cells = of_n_size_cells(memory);
259 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
261 unsigned long result = 0;
264 result = (result << 32) | **buf;
271 * Figure out to which domain a cpu belongs and stick it there.
272 * Return the id of the domain used.
274 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
277 struct device_node *cpu = find_cpu_node(lcpu);
284 nid = of_node_to_nid_single(cpu);
286 if (nid < 0 || !node_online(nid))
287 nid = any_online_node(NODE_MASK_ALL);
289 map_cpu_to_node(lcpu, nid);
296 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
297 unsigned long action,
300 unsigned long lcpu = (unsigned long)hcpu;
301 int ret = NOTIFY_DONE;
305 case CPU_UP_PREPARE_FROZEN:
306 numa_setup_cpu(lcpu);
309 #ifdef CONFIG_HOTPLUG_CPU
311 case CPU_DEAD_FROZEN:
312 case CPU_UP_CANCELED:
313 case CPU_UP_CANCELED_FROZEN:
314 unmap_cpu_from_node(lcpu);
323 * Check and possibly modify a memory region to enforce the memory limit.
325 * Returns the size the region should have to enforce the memory limit.
326 * This will either be the original value of size, a truncated value,
327 * or zero. If the returned value of size is 0 the region should be
328 * discarded as it lies wholy above the memory limit.
330 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
334 * We use lmb_end_of_DRAM() in here instead of memory_limit because
335 * we've already adjusted it for the limit and it takes care of
336 * having memory holes below the limit.
342 if (start + size <= lmb_end_of_DRAM())
345 if (start >= lmb_end_of_DRAM())
348 return lmb_end_of_DRAM() - start;
352 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
353 * node. This assumes n_mem_{addr,size}_cells have been set.
355 static void __init parse_drconf_memory(struct device_node *memory)
357 const unsigned int *lm, *dm, *aa;
358 unsigned int ls, ld, la;
359 unsigned int n, aam, aalen;
360 unsigned long lmb_size, size, start;
361 int nid, default_nid = 0;
362 unsigned int ai, flags;
364 lm = of_get_property(memory, "ibm,lmb-size", &ls);
365 dm = of_get_property(memory, "ibm,dynamic-memory", &ld);
366 aa = of_get_property(memory, "ibm,associativity-lookup-arrays", &la);
367 if (!lm || !dm || !aa ||
368 ls < sizeof(unsigned int) || ld < sizeof(unsigned int) ||
369 la < 2 * sizeof(unsigned int))
372 lmb_size = read_n_cells(n_mem_size_cells, &lm);
373 n = *dm++; /* number of LMBs */
374 aam = *aa++; /* number of associativity lists */
375 aalen = *aa++; /* length of each associativity list */
376 if (ld < (n * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int) ||
377 la < (aam * aalen + 2) * sizeof(unsigned int))
380 for (; n != 0; --n) {
381 start = read_n_cells(n_mem_addr_cells, &dm);
385 /* 0x80 == reserved, 0x8 = assigned to us */
386 if ((flags & 0x80) || !(flags & 0x8))
389 /* flags & 0x40 means associativity index is invalid */
390 if (min_common_depth > 0 && min_common_depth <= aalen &&
391 (flags & 0x40) == 0 && ai < aam) {
392 /* this is like of_node_to_nid_single */
393 nid = aa[ai * aalen + min_common_depth - 1];
394 if (nid == 0xffff || nid >= MAX_NUMNODES)
398 fake_numa_create_new_node(((start + lmb_size) >> PAGE_SHIFT),
400 node_set_online(nid);
402 size = numa_enforce_memory_limit(start, lmb_size);
406 add_active_range(nid, start >> PAGE_SHIFT,
407 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
411 static int __init parse_numa_properties(void)
413 struct device_node *cpu = NULL;
414 struct device_node *memory = NULL;
418 if (numa_enabled == 0) {
419 printk(KERN_WARNING "NUMA disabled by user\n");
423 min_common_depth = find_min_common_depth();
425 if (min_common_depth < 0)
426 return min_common_depth;
428 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
431 * Even though we connect cpus to numa domains later in SMP
432 * init, we need to know the node ids now. This is because
433 * each node to be onlined must have NODE_DATA etc backing it.
435 for_each_present_cpu(i) {
438 cpu = find_cpu_node(i);
440 nid = of_node_to_nid_single(cpu);
444 * Don't fall back to default_nid yet -- we will plug
445 * cpus into nodes once the memory scan has discovered
450 node_set_online(nid);
453 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
455 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
460 const unsigned int *memcell_buf;
463 memcell_buf = of_get_property(memory,
464 "linux,usable-memory", &len);
465 if (!memcell_buf || len <= 0)
466 memcell_buf = of_get_property(memory, "reg", &len);
467 if (!memcell_buf || len <= 0)
471 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
473 /* these are order-sensitive, and modify the buffer pointer */
474 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
475 size = read_n_cells(n_mem_size_cells, &memcell_buf);
478 * Assumption: either all memory nodes or none will
479 * have associativity properties. If none, then
480 * everything goes to default_nid.
482 nid = of_node_to_nid_single(memory);
486 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
487 node_set_online(nid);
489 if (!(size = numa_enforce_memory_limit(start, size))) {
496 add_active_range(nid, start >> PAGE_SHIFT,
497 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
504 * Now do the same thing for each LMB listed in the ibm,dynamic-memory
505 * property in the ibm,dynamic-reconfiguration-memory node.
507 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
509 parse_drconf_memory(memory);
514 static void __init setup_nonnuma(void)
516 unsigned long top_of_ram = lmb_end_of_DRAM();
517 unsigned long total_ram = lmb_phys_mem_size();
518 unsigned long start_pfn, end_pfn;
519 unsigned int i, nid = 0;
521 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
522 top_of_ram, total_ram);
523 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
524 (top_of_ram - total_ram) >> 20);
526 for (i = 0; i < lmb.memory.cnt; ++i) {
527 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
528 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
530 fake_numa_create_new_node(end_pfn, &nid);
531 add_active_range(nid, start_pfn, end_pfn);
532 node_set_online(nid);
536 void __init dump_numa_cpu_topology(void)
539 unsigned int cpu, count;
541 if (min_common_depth == -1 || !numa_enabled)
544 for_each_online_node(node) {
545 printk(KERN_DEBUG "Node %d CPUs:", node);
549 * If we used a CPU iterator here we would miss printing
550 * the holes in the cpumap.
552 for (cpu = 0; cpu < NR_CPUS; cpu++) {
553 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
559 printk("-%u", cpu - 1);
565 printk("-%u", NR_CPUS - 1);
570 static void __init dump_numa_memory_topology(void)
575 if (min_common_depth == -1 || !numa_enabled)
578 for_each_online_node(node) {
581 printk(KERN_DEBUG "Node %d Memory:", node);
585 for (i = 0; i < lmb_end_of_DRAM();
586 i += (1 << SECTION_SIZE_BITS)) {
587 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
605 * Allocate some memory, satisfying the lmb or bootmem allocator where
606 * required. nid is the preferred node and end is the physical address of
607 * the highest address in the node.
609 * Returns the physical address of the memory.
611 static void __init *careful_allocation(int nid, unsigned long size,
613 unsigned long end_pfn)
616 unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
618 /* retry over all memory */
620 ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
623 panic("numa.c: cannot allocate %lu bytes on node %d",
627 * If the memory came from a previously allocated node, we must
628 * retry with the bootmem allocator.
630 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
632 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
636 panic("numa.c: cannot allocate %lu bytes on node %d",
641 dbg("alloc_bootmem %lx %lx\n", ret, size);
647 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
648 .notifier_call = cpu_numa_callback,
649 .priority = 1 /* Must run before sched domains notifier. */
652 void __init do_init_bootmem(void)
658 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
659 max_pfn = max_low_pfn;
661 if (parse_numa_properties())
664 dump_numa_memory_topology();
666 register_cpu_notifier(&ppc64_numa_nb);
667 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
668 (void *)(unsigned long)boot_cpuid);
670 for_each_online_node(nid) {
671 unsigned long start_pfn, end_pfn;
672 unsigned long bootmem_paddr;
673 unsigned long bootmap_pages;
675 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
677 /* Allocate the node structure node local if possible */
678 NODE_DATA(nid) = careful_allocation(nid,
679 sizeof(struct pglist_data),
680 SMP_CACHE_BYTES, end_pfn);
681 NODE_DATA(nid) = __va(NODE_DATA(nid));
682 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
684 dbg("node %d\n", nid);
685 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
687 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
688 NODE_DATA(nid)->node_start_pfn = start_pfn;
689 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
691 if (NODE_DATA(nid)->node_spanned_pages == 0)
694 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
695 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
697 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
698 bootmem_paddr = (unsigned long)careful_allocation(nid,
699 bootmap_pages << PAGE_SHIFT,
701 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
703 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
705 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
708 free_bootmem_with_active_regions(nid, end_pfn);
710 /* Mark reserved regions on this node */
711 for (i = 0; i < lmb.reserved.cnt; i++) {
712 unsigned long physbase = lmb.reserved.region[i].base;
713 unsigned long size = lmb.reserved.region[i].size;
714 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
715 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
717 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
718 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
721 if (physbase < end_paddr &&
722 (physbase+size) > start_paddr) {
724 if (physbase < start_paddr) {
725 size -= start_paddr - physbase;
726 physbase = start_paddr;
729 if (size > end_paddr - physbase)
730 size = end_paddr - physbase;
732 dbg("reserve_bootmem %lx %lx\n", physbase,
734 reserve_bootmem_node(NODE_DATA(nid), physbase,
735 size, BOOTMEM_DEFAULT);
739 sparse_memory_present_with_active_regions(nid);
743 void __init paging_init(void)
745 unsigned long max_zone_pfns[MAX_NR_ZONES];
746 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
747 max_zone_pfns[ZONE_DMA] = lmb_end_of_DRAM() >> PAGE_SHIFT;
748 free_area_init_nodes(max_zone_pfns);
751 static int __init early_numa(char *p)
756 if (strstr(p, "off"))
759 if (strstr(p, "debug"))
762 p = strstr(p, "fake=");
764 cmdline = p + strlen("fake=");
768 early_param("numa", early_numa);
770 #ifdef CONFIG_MEMORY_HOTPLUG
772 * Find the node associated with a hot added memory section. Section
773 * corresponds to a SPARSEMEM section, not an LMB. It is assumed that
774 * sections are fully contained within a single LMB.
776 int hot_add_scn_to_nid(unsigned long scn_addr)
778 struct device_node *memory = NULL;
780 int default_nid = any_online_node(NODE_MASK_ALL);
783 if (!numa_enabled || (min_common_depth < 0))
786 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
787 unsigned long start, size;
789 const unsigned int *memcell_buf;
792 memcell_buf = of_get_property(memory, "reg", &len);
793 if (!memcell_buf || len <= 0)
797 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
799 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
800 size = read_n_cells(n_mem_size_cells, &memcell_buf);
801 nid = of_node_to_nid_single(memory);
803 /* Domains not present at boot default to 0 */
804 if (nid < 0 || !node_online(nid))
807 if ((scn_addr >= start) && (scn_addr < (start + size))) {
812 if (--ranges) /* process all ranges in cell */
815 BUG(); /* section address should be found above */
818 /* Temporary code to ensure that returned node is not empty */
821 while (NODE_DATA(nid)->node_spanned_pages == 0) {
822 node_clear(nid, nodes);
823 nid = any_online_node(nodes);
827 #endif /* CONFIG_MEMORY_HOTPLUG */