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 char *cmdline __initdata;
29 static int numa_debug;
30 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
32 int numa_cpu_lookup_table[NR_CPUS];
33 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
34 struct pglist_data *node_data[MAX_NUMNODES];
36 EXPORT_SYMBOL(numa_cpu_lookup_table);
37 EXPORT_SYMBOL(numa_cpumask_lookup_table);
38 EXPORT_SYMBOL(node_data);
40 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
41 static int min_common_depth;
42 static int n_mem_addr_cells, n_mem_size_cells;
44 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
47 unsigned long long mem;
49 static unsigned int fake_nid = 0;
50 static unsigned long long curr_boundary = 0;
56 mem = memparse(p, &p);
60 if (mem < curr_boundary)
65 if ((end_pfn << PAGE_SHIFT) > mem) {
67 * Skip commas and spaces
69 while (*p == ',' || *p == ' ' || *p == '\t')
75 dbg("created new fake_node with id %d\n", fake_nid);
81 static void __cpuinit map_cpu_to_node(int cpu, int node)
83 numa_cpu_lookup_table[cpu] = node;
85 dbg("adding cpu %d to node %d\n", cpu, node);
87 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
88 cpu_set(cpu, numa_cpumask_lookup_table[node]);
91 #ifdef CONFIG_HOTPLUG_CPU
92 static void unmap_cpu_from_node(unsigned long cpu)
94 int node = numa_cpu_lookup_table[cpu];
96 dbg("removing cpu %lu from node %d\n", cpu, node);
98 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
99 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
101 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
105 #endif /* CONFIG_HOTPLUG_CPU */
107 static struct device_node * __cpuinit find_cpu_node(unsigned int cpu)
109 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
110 struct device_node *cpu_node = NULL;
111 const unsigned int *interrupt_server, *reg;
114 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
115 /* Try interrupt server first */
116 interrupt_server = of_get_property(cpu_node,
117 "ibm,ppc-interrupt-server#s", &len);
119 len = len / sizeof(u32);
121 if (interrupt_server && (len > 0)) {
123 if (interrupt_server[len] == hw_cpuid)
127 reg = of_get_property(cpu_node, "reg", &len);
128 if (reg && (len > 0) && (reg[0] == hw_cpuid))
136 /* must hold reference to node during call */
137 static const int *of_get_associativity(struct device_node *dev)
139 return of_get_property(dev, "ibm,associativity", NULL);
142 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
145 static int of_node_to_nid_single(struct device_node *device)
148 const unsigned int *tmp;
150 if (min_common_depth == -1)
153 tmp = of_get_associativity(device);
157 if (tmp[0] >= min_common_depth)
158 nid = tmp[min_common_depth];
160 /* POWER4 LPAR uses 0xffff as invalid node */
161 if (nid == 0xffff || nid >= MAX_NUMNODES)
167 /* Walk the device tree upwards, looking for an associativity id */
168 int of_node_to_nid(struct device_node *device)
170 struct device_node *tmp;
175 nid = of_node_to_nid_single(device);
180 device = of_get_parent(tmp);
187 EXPORT_SYMBOL_GPL(of_node_to_nid);
190 * In theory, the "ibm,associativity" property may contain multiple
191 * associativity lists because a resource may be multiply connected
192 * into the machine. This resource then has different associativity
193 * characteristics relative to its multiple connections. We ignore
194 * this for now. We also assume that all cpu and memory sets have
195 * their distances represented at a common level. This won't be
196 * true for hierarchical NUMA.
198 * In any case the ibm,associativity-reference-points should give
199 * the correct depth for a normal NUMA system.
201 * - Dave Hansen <haveblue@us.ibm.com>
203 static int __init find_min_common_depth(void)
206 const unsigned int *ref_points;
207 struct device_node *rtas_root;
210 rtas_root = of_find_node_by_path("/rtas");
216 * this property is 2 32-bit integers, each representing a level of
217 * depth in the associativity nodes. The first is for an SMP
218 * configuration (should be all 0's) and the second is for a normal
219 * NUMA configuration.
221 ref_points = of_get_property(rtas_root,
222 "ibm,associativity-reference-points", &len);
224 if ((len >= 1) && ref_points) {
225 depth = ref_points[1];
227 dbg("NUMA: ibm,associativity-reference-points not found.\n");
230 of_node_put(rtas_root);
235 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
237 struct device_node *memory = NULL;
239 memory = of_find_node_by_type(memory, "memory");
241 panic("numa.c: No memory nodes found!");
243 *n_addr_cells = of_n_addr_cells(memory);
244 *n_size_cells = of_n_size_cells(memory);
248 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
250 unsigned long result = 0;
253 result = (result << 32) | **buf;
260 * Figure out to which domain a cpu belongs and stick it there.
261 * Return the id of the domain used.
263 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
266 struct device_node *cpu = find_cpu_node(lcpu);
273 nid = of_node_to_nid_single(cpu);
275 if (nid < 0 || !node_online(nid))
276 nid = any_online_node(NODE_MASK_ALL);
278 map_cpu_to_node(lcpu, nid);
285 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
286 unsigned long action,
289 unsigned long lcpu = (unsigned long)hcpu;
290 int ret = NOTIFY_DONE;
294 case CPU_UP_PREPARE_FROZEN:
295 numa_setup_cpu(lcpu);
298 #ifdef CONFIG_HOTPLUG_CPU
300 case CPU_DEAD_FROZEN:
301 case CPU_UP_CANCELED:
302 case CPU_UP_CANCELED_FROZEN:
303 unmap_cpu_from_node(lcpu);
312 * Check and possibly modify a memory region to enforce the memory limit.
314 * Returns the size the region should have to enforce the memory limit.
315 * This will either be the original value of size, a truncated value,
316 * or zero. If the returned value of size is 0 the region should be
317 * discarded as it lies wholy above the memory limit.
319 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
323 * We use lmb_end_of_DRAM() in here instead of memory_limit because
324 * we've already adjusted it for the limit and it takes care of
325 * having memory holes below the limit.
331 if (start + size <= lmb_end_of_DRAM())
334 if (start >= lmb_end_of_DRAM())
337 return lmb_end_of_DRAM() - start;
341 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
342 * node. This assumes n_mem_{addr,size}_cells have been set.
344 static void __init parse_drconf_memory(struct device_node *memory)
346 const unsigned int *lm, *dm, *aa;
347 unsigned int ls, ld, la;
348 unsigned int n, aam, aalen;
349 unsigned long lmb_size, size, start;
350 int nid, default_nid = 0;
351 unsigned int ai, flags;
353 lm = of_get_property(memory, "ibm,lmb-size", &ls);
354 dm = of_get_property(memory, "ibm,dynamic-memory", &ld);
355 aa = of_get_property(memory, "ibm,associativity-lookup-arrays", &la);
356 if (!lm || !dm || !aa ||
357 ls < sizeof(unsigned int) || ld < sizeof(unsigned int) ||
358 la < 2 * sizeof(unsigned int))
361 lmb_size = read_n_cells(n_mem_size_cells, &lm);
362 n = *dm++; /* number of LMBs */
363 aam = *aa++; /* number of associativity lists */
364 aalen = *aa++; /* length of each associativity list */
365 if (ld < (n * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int) ||
366 la < (aam * aalen + 2) * sizeof(unsigned int))
369 for (; n != 0; --n) {
370 start = read_n_cells(n_mem_addr_cells, &dm);
374 /* 0x80 == reserved, 0x8 = assigned to us */
375 if ((flags & 0x80) || !(flags & 0x8))
378 /* flags & 0x40 means associativity index is invalid */
379 if (min_common_depth > 0 && min_common_depth <= aalen &&
380 (flags & 0x40) == 0 && ai < aam) {
381 /* this is like of_node_to_nid_single */
382 nid = aa[ai * aalen + min_common_depth - 1];
383 if (nid == 0xffff || nid >= MAX_NUMNODES)
387 size = numa_enforce_memory_limit(start, lmb_size);
391 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
392 node_set_online(nid);
394 add_active_range(nid, start >> PAGE_SHIFT,
395 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
399 static int __init parse_numa_properties(void)
401 struct device_node *cpu = NULL;
402 struct device_node *memory = NULL;
406 if (numa_enabled == 0) {
407 printk(KERN_WARNING "NUMA disabled by user\n");
411 min_common_depth = find_min_common_depth();
413 if (min_common_depth < 0)
414 return min_common_depth;
416 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
419 * Even though we connect cpus to numa domains later in SMP
420 * init, we need to know the node ids now. This is because
421 * each node to be onlined must have NODE_DATA etc backing it.
423 for_each_present_cpu(i) {
426 cpu = find_cpu_node(i);
428 nid = of_node_to_nid_single(cpu);
432 * Don't fall back to default_nid yet -- we will plug
433 * cpus into nodes once the memory scan has discovered
438 node_set_online(nid);
441 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
443 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
448 const unsigned int *memcell_buf;
451 memcell_buf = of_get_property(memory,
452 "linux,usable-memory", &len);
453 if (!memcell_buf || len <= 0)
454 memcell_buf = of_get_property(memory, "reg", &len);
455 if (!memcell_buf || len <= 0)
459 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
461 /* these are order-sensitive, and modify the buffer pointer */
462 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
463 size = read_n_cells(n_mem_size_cells, &memcell_buf);
466 * Assumption: either all memory nodes or none will
467 * have associativity properties. If none, then
468 * everything goes to default_nid.
470 nid = of_node_to_nid_single(memory);
474 if (!(size = numa_enforce_memory_limit(start, size))) {
481 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
482 node_set_online(nid);
484 add_active_range(nid, start >> PAGE_SHIFT,
485 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
492 * Now do the same thing for each LMB listed in the ibm,dynamic-memory
493 * property in the ibm,dynamic-reconfiguration-memory node.
495 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
497 parse_drconf_memory(memory);
502 static void __init setup_nonnuma(void)
504 unsigned long top_of_ram = lmb_end_of_DRAM();
505 unsigned long total_ram = lmb_phys_mem_size();
506 unsigned long start_pfn, end_pfn;
507 unsigned int i, nid = 0;
509 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
510 top_of_ram, total_ram);
511 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
512 (top_of_ram - total_ram) >> 20);
514 for (i = 0; i < lmb.memory.cnt; ++i) {
515 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
516 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
518 fake_numa_create_new_node(end_pfn, &nid);
519 add_active_range(nid, start_pfn, end_pfn);
520 node_set_online(nid);
524 void __init dump_numa_cpu_topology(void)
527 unsigned int cpu, count;
529 if (min_common_depth == -1 || !numa_enabled)
532 for_each_online_node(node) {
533 printk(KERN_DEBUG "Node %d CPUs:", node);
537 * If we used a CPU iterator here we would miss printing
538 * the holes in the cpumap.
540 for (cpu = 0; cpu < NR_CPUS; cpu++) {
541 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
547 printk("-%u", cpu - 1);
553 printk("-%u", NR_CPUS - 1);
558 static void __init dump_numa_memory_topology(void)
563 if (min_common_depth == -1 || !numa_enabled)
566 for_each_online_node(node) {
569 printk(KERN_DEBUG "Node %d Memory:", node);
573 for (i = 0; i < lmb_end_of_DRAM();
574 i += (1 << SECTION_SIZE_BITS)) {
575 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
593 * Allocate some memory, satisfying the lmb or bootmem allocator where
594 * required. nid is the preferred node and end is the physical address of
595 * the highest address in the node.
597 * Returns the physical address of the memory.
599 static void __init *careful_allocation(int nid, unsigned long size,
601 unsigned long end_pfn)
604 unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
606 /* retry over all memory */
608 ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
611 panic("numa.c: cannot allocate %lu bytes on node %d",
615 * If the memory came from a previously allocated node, we must
616 * retry with the bootmem allocator.
618 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
620 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
624 panic("numa.c: cannot allocate %lu bytes on node %d",
629 dbg("alloc_bootmem %lx %lx\n", ret, size);
635 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
636 .notifier_call = cpu_numa_callback,
637 .priority = 1 /* Must run before sched domains notifier. */
640 void __init do_init_bootmem(void)
646 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
647 max_pfn = max_low_pfn;
649 if (parse_numa_properties())
652 dump_numa_memory_topology();
654 register_cpu_notifier(&ppc64_numa_nb);
655 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
656 (void *)(unsigned long)boot_cpuid);
658 for_each_online_node(nid) {
659 unsigned long start_pfn, end_pfn;
660 unsigned long bootmem_paddr;
661 unsigned long bootmap_pages;
663 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
665 /* Allocate the node structure node local if possible */
666 NODE_DATA(nid) = careful_allocation(nid,
667 sizeof(struct pglist_data),
668 SMP_CACHE_BYTES, end_pfn);
669 NODE_DATA(nid) = __va(NODE_DATA(nid));
670 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
672 dbg("node %d\n", nid);
673 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
675 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
676 NODE_DATA(nid)->node_start_pfn = start_pfn;
677 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
679 if (NODE_DATA(nid)->node_spanned_pages == 0)
682 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
683 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
685 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
686 bootmem_paddr = (unsigned long)careful_allocation(nid,
687 bootmap_pages << PAGE_SHIFT,
689 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
691 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
693 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
696 free_bootmem_with_active_regions(nid, end_pfn);
698 /* Mark reserved regions on this node */
699 for (i = 0; i < lmb.reserved.cnt; i++) {
700 unsigned long physbase = lmb.reserved.region[i].base;
701 unsigned long size = lmb.reserved.region[i].size;
702 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
703 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
705 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
706 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
709 if (physbase < end_paddr &&
710 (physbase+size) > start_paddr) {
712 if (physbase < start_paddr) {
713 size -= start_paddr - physbase;
714 physbase = start_paddr;
717 if (size > end_paddr - physbase)
718 size = end_paddr - physbase;
720 dbg("reserve_bootmem %lx %lx\n", physbase,
722 reserve_bootmem_node(NODE_DATA(nid), physbase,
727 sparse_memory_present_with_active_regions(nid);
731 void __init paging_init(void)
733 unsigned long max_zone_pfns[MAX_NR_ZONES];
734 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
735 max_zone_pfns[ZONE_DMA] = lmb_end_of_DRAM() >> PAGE_SHIFT;
736 free_area_init_nodes(max_zone_pfns);
739 static int __init early_numa(char *p)
744 if (strstr(p, "off"))
747 if (strstr(p, "debug"))
750 p = strstr(p, "fake=");
752 cmdline = p + strlen("fake=");
756 early_param("numa", early_numa);
758 #ifdef CONFIG_MEMORY_HOTPLUG
760 * Find the node associated with a hot added memory section. Section
761 * corresponds to a SPARSEMEM section, not an LMB. It is assumed that
762 * sections are fully contained within a single LMB.
764 int hot_add_scn_to_nid(unsigned long scn_addr)
766 struct device_node *memory = NULL;
768 int default_nid = any_online_node(NODE_MASK_ALL);
771 if (!numa_enabled || (min_common_depth < 0))
774 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
775 unsigned long start, size;
777 const unsigned int *memcell_buf;
780 memcell_buf = of_get_property(memory, "reg", &len);
781 if (!memcell_buf || len <= 0)
785 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
787 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
788 size = read_n_cells(n_mem_size_cells, &memcell_buf);
789 nid = of_node_to_nid_single(memory);
791 /* Domains not present at boot default to 0 */
792 if (nid < 0 || !node_online(nid))
795 if ((scn_addr >= start) && (scn_addr < (start + size))) {
800 if (--ranges) /* process all ranges in cell */
803 BUG(); /* section address should be found above */
806 /* Temporary code to ensure that returned node is not empty */
809 while (NODE_DATA(nid)->node_spanned_pages == 0) {
810 node_clear(nid, nodes);
811 nid = any_online_node(nodes);
815 #endif /* CONFIG_MEMORY_HOTPLUG */