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>
22 #include <asm/sparsemem.h>
24 #include <asm/system.h>
27 static int numa_enabled = 1;
29 static char *cmdline __initdata;
31 static int numa_debug;
32 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
34 int numa_cpu_lookup_table[NR_CPUS];
35 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
36 struct pglist_data *node_data[MAX_NUMNODES];
38 EXPORT_SYMBOL(numa_cpu_lookup_table);
39 EXPORT_SYMBOL(numa_cpumask_lookup_table);
40 EXPORT_SYMBOL(node_data);
42 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
43 static int min_common_depth;
44 static int n_mem_addr_cells, n_mem_size_cells;
46 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
49 unsigned long long mem;
51 static unsigned int fake_nid;
52 static unsigned long long curr_boundary;
55 * Modify node id, iff we started creating NUMA nodes
56 * We want to continue from where we left of the last time
61 * In case there are no more arguments to parse, the
62 * node_id should be the same as the last fake node id
63 * (we've handled this above).
68 mem = memparse(p, &p);
72 if (mem < curr_boundary)
77 if ((end_pfn << PAGE_SHIFT) > mem) {
79 * Skip commas and spaces
81 while (*p == ',' || *p == ' ' || *p == '\t')
87 dbg("created new fake_node with id %d\n", fake_nid);
93 static void __cpuinit map_cpu_to_node(int cpu, int node)
95 numa_cpu_lookup_table[cpu] = node;
97 dbg("adding cpu %d to node %d\n", cpu, node);
99 if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
100 cpu_set(cpu, numa_cpumask_lookup_table[node]);
103 #ifdef CONFIG_HOTPLUG_CPU
104 static void unmap_cpu_from_node(unsigned long cpu)
106 int node = numa_cpu_lookup_table[cpu];
108 dbg("removing cpu %lu from node %d\n", cpu, node);
110 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
111 cpu_clear(cpu, numa_cpumask_lookup_table[node]);
113 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
117 #endif /* CONFIG_HOTPLUG_CPU */
119 static struct device_node * __cpuinit find_cpu_node(unsigned int cpu)
121 unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
122 struct device_node *cpu_node = NULL;
123 const unsigned int *interrupt_server, *reg;
126 while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
127 /* Try interrupt server first */
128 interrupt_server = of_get_property(cpu_node,
129 "ibm,ppc-interrupt-server#s", &len);
131 len = len / sizeof(u32);
133 if (interrupt_server && (len > 0)) {
135 if (interrupt_server[len] == hw_cpuid)
139 reg = of_get_property(cpu_node, "reg", &len);
140 if (reg && (len > 0) && (reg[0] == hw_cpuid))
148 /* must hold reference to node during call */
149 static const int *of_get_associativity(struct device_node *dev)
151 return of_get_property(dev, "ibm,associativity", NULL);
154 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
157 static int of_node_to_nid_single(struct device_node *device)
160 const unsigned int *tmp;
162 if (min_common_depth == -1)
165 tmp = of_get_associativity(device);
169 if (tmp[0] >= min_common_depth)
170 nid = tmp[min_common_depth];
172 /* POWER4 LPAR uses 0xffff as invalid node */
173 if (nid == 0xffff || nid >= MAX_NUMNODES)
179 /* Walk the device tree upwards, looking for an associativity id */
180 int of_node_to_nid(struct device_node *device)
182 struct device_node *tmp;
187 nid = of_node_to_nid_single(device);
192 device = of_get_parent(tmp);
199 EXPORT_SYMBOL_GPL(of_node_to_nid);
202 * In theory, the "ibm,associativity" property may contain multiple
203 * associativity lists because a resource may be multiply connected
204 * into the machine. This resource then has different associativity
205 * characteristics relative to its multiple connections. We ignore
206 * this for now. We also assume that all cpu and memory sets have
207 * their distances represented at a common level. This won't be
208 * true for hierarchical NUMA.
210 * In any case the ibm,associativity-reference-points should give
211 * the correct depth for a normal NUMA system.
213 * - Dave Hansen <haveblue@us.ibm.com>
215 static int __init find_min_common_depth(void)
218 const unsigned int *ref_points;
219 struct device_node *rtas_root;
222 rtas_root = of_find_node_by_path("/rtas");
228 * this property is 2 32-bit integers, each representing a level of
229 * depth in the associativity nodes. The first is for an SMP
230 * configuration (should be all 0's) and the second is for a normal
231 * NUMA configuration.
233 ref_points = of_get_property(rtas_root,
234 "ibm,associativity-reference-points", &len);
236 if ((len >= 1) && ref_points) {
237 depth = ref_points[1];
239 dbg("NUMA: ibm,associativity-reference-points not found.\n");
242 of_node_put(rtas_root);
247 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
249 struct device_node *memory = NULL;
251 memory = of_find_node_by_type(memory, "memory");
253 panic("numa.c: No memory nodes found!");
255 *n_addr_cells = of_n_addr_cells(memory);
256 *n_size_cells = of_n_size_cells(memory);
260 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
262 unsigned long result = 0;
265 result = (result << 32) | **buf;
272 * Figure out to which domain a cpu belongs and stick it there.
273 * Return the id of the domain used.
275 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
278 struct device_node *cpu = find_cpu_node(lcpu);
285 nid = of_node_to_nid_single(cpu);
287 if (nid < 0 || !node_online(nid))
288 nid = any_online_node(NODE_MASK_ALL);
290 map_cpu_to_node(lcpu, nid);
297 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
298 unsigned long action,
301 unsigned long lcpu = (unsigned long)hcpu;
302 int ret = NOTIFY_DONE;
306 case CPU_UP_PREPARE_FROZEN:
307 numa_setup_cpu(lcpu);
310 #ifdef CONFIG_HOTPLUG_CPU
312 case CPU_DEAD_FROZEN:
313 case CPU_UP_CANCELED:
314 case CPU_UP_CANCELED_FROZEN:
315 unmap_cpu_from_node(lcpu);
324 * Check and possibly modify a memory region to enforce the memory limit.
326 * Returns the size the region should have to enforce the memory limit.
327 * This will either be the original value of size, a truncated value,
328 * or zero. If the returned value of size is 0 the region should be
329 * discarded as it lies wholy above the memory limit.
331 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
335 * We use lmb_end_of_DRAM() in here instead of memory_limit because
336 * we've already adjusted it for the limit and it takes care of
337 * having memory holes below the limit.
343 if (start + size <= lmb_end_of_DRAM())
346 if (start >= lmb_end_of_DRAM())
349 return lmb_end_of_DRAM() - start;
353 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
354 * node. This assumes n_mem_{addr,size}_cells have been set.
356 static void __init parse_drconf_memory(struct device_node *memory)
358 const unsigned int *lm, *dm, *aa;
359 unsigned int ls, ld, la;
360 unsigned int n, aam, aalen;
361 unsigned long lmb_size, size, start;
362 int nid, default_nid = 0;
363 unsigned int ai, flags;
365 lm = of_get_property(memory, "ibm,lmb-size", &ls);
366 dm = of_get_property(memory, "ibm,dynamic-memory", &ld);
367 aa = of_get_property(memory, "ibm,associativity-lookup-arrays", &la);
368 if (!lm || !dm || !aa ||
369 ls < sizeof(unsigned int) || ld < sizeof(unsigned int) ||
370 la < 2 * sizeof(unsigned int))
373 lmb_size = read_n_cells(n_mem_size_cells, &lm);
374 n = *dm++; /* number of LMBs */
375 aam = *aa++; /* number of associativity lists */
376 aalen = *aa++; /* length of each associativity list */
377 if (ld < (n * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int) ||
378 la < (aam * aalen + 2) * sizeof(unsigned int))
381 for (; n != 0; --n) {
382 start = read_n_cells(n_mem_addr_cells, &dm);
386 /* 0x80 == reserved, 0x8 = assigned to us */
387 if ((flags & 0x80) || !(flags & 0x8))
390 /* flags & 0x40 means associativity index is invalid */
391 if (min_common_depth > 0 && min_common_depth <= aalen &&
392 (flags & 0x40) == 0 && ai < aam) {
393 /* this is like of_node_to_nid_single */
394 nid = aa[ai * aalen + min_common_depth - 1];
395 if (nid == 0xffff || nid >= MAX_NUMNODES)
399 fake_numa_create_new_node(((start + lmb_size) >> PAGE_SHIFT),
401 node_set_online(nid);
403 size = numa_enforce_memory_limit(start, lmb_size);
407 add_active_range(nid, start >> PAGE_SHIFT,
408 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
412 static int __init parse_numa_properties(void)
414 struct device_node *cpu = NULL;
415 struct device_node *memory = NULL;
419 if (numa_enabled == 0) {
420 printk(KERN_WARNING "NUMA disabled by user\n");
424 min_common_depth = find_min_common_depth();
426 if (min_common_depth < 0)
427 return min_common_depth;
429 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
432 * Even though we connect cpus to numa domains later in SMP
433 * init, we need to know the node ids now. This is because
434 * each node to be onlined must have NODE_DATA etc backing it.
436 for_each_present_cpu(i) {
439 cpu = find_cpu_node(i);
441 nid = of_node_to_nid_single(cpu);
445 * Don't fall back to default_nid yet -- we will plug
446 * cpus into nodes once the memory scan has discovered
451 node_set_online(nid);
454 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
456 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
461 const unsigned int *memcell_buf;
464 memcell_buf = of_get_property(memory,
465 "linux,usable-memory", &len);
466 if (!memcell_buf || len <= 0)
467 memcell_buf = of_get_property(memory, "reg", &len);
468 if (!memcell_buf || len <= 0)
472 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
474 /* these are order-sensitive, and modify the buffer pointer */
475 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
476 size = read_n_cells(n_mem_size_cells, &memcell_buf);
479 * Assumption: either all memory nodes or none will
480 * have associativity properties. If none, then
481 * everything goes to default_nid.
483 nid = of_node_to_nid_single(memory);
487 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
488 node_set_online(nid);
490 if (!(size = numa_enforce_memory_limit(start, size))) {
497 add_active_range(nid, start >> PAGE_SHIFT,
498 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
505 * Now do the same thing for each LMB listed in the ibm,dynamic-memory
506 * property in the ibm,dynamic-reconfiguration-memory node.
508 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
510 parse_drconf_memory(memory);
515 static void __init setup_nonnuma(void)
517 unsigned long top_of_ram = lmb_end_of_DRAM();
518 unsigned long total_ram = lmb_phys_mem_size();
519 unsigned long start_pfn, end_pfn;
520 unsigned int i, nid = 0;
522 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
523 top_of_ram, total_ram);
524 printk(KERN_DEBUG "Memory hole size: %ldMB\n",
525 (top_of_ram - total_ram) >> 20);
527 for (i = 0; i < lmb.memory.cnt; ++i) {
528 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
529 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
531 fake_numa_create_new_node(end_pfn, &nid);
532 add_active_range(nid, start_pfn, end_pfn);
533 node_set_online(nid);
537 void __init dump_numa_cpu_topology(void)
540 unsigned int cpu, count;
542 if (min_common_depth == -1 || !numa_enabled)
545 for_each_online_node(node) {
546 printk(KERN_DEBUG "Node %d CPUs:", node);
550 * If we used a CPU iterator here we would miss printing
551 * the holes in the cpumap.
553 for (cpu = 0; cpu < NR_CPUS; cpu++) {
554 if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
560 printk("-%u", cpu - 1);
566 printk("-%u", NR_CPUS - 1);
571 static void __init dump_numa_memory_topology(void)
576 if (min_common_depth == -1 || !numa_enabled)
579 for_each_online_node(node) {
582 printk(KERN_DEBUG "Node %d Memory:", node);
586 for (i = 0; i < lmb_end_of_DRAM();
587 i += (1 << SECTION_SIZE_BITS)) {
588 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
606 * Allocate some memory, satisfying the lmb or bootmem allocator where
607 * required. nid is the preferred node and end is the physical address of
608 * the highest address in the node.
610 * Returns the physical address of the memory.
612 static void __init *careful_allocation(int nid, unsigned long size,
614 unsigned long end_pfn)
617 unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
619 /* retry over all memory */
621 ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
624 panic("numa.c: cannot allocate %lu bytes on node %d",
628 * If the memory came from a previously allocated node, we must
629 * retry with the bootmem allocator.
631 new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
633 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
637 panic("numa.c: cannot allocate %lu bytes on node %d",
642 dbg("alloc_bootmem %lx %lx\n", ret, size);
648 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
649 .notifier_call = cpu_numa_callback,
650 .priority = 1 /* Must run before sched domains notifier. */
653 void __init do_init_bootmem(void)
659 max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
660 max_pfn = max_low_pfn;
662 if (parse_numa_properties())
665 dump_numa_memory_topology();
667 register_cpu_notifier(&ppc64_numa_nb);
668 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
669 (void *)(unsigned long)boot_cpuid);
671 for_each_online_node(nid) {
672 unsigned long start_pfn, end_pfn;
673 unsigned long bootmem_paddr;
674 unsigned long bootmap_pages;
676 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
678 /* Allocate the node structure node local if possible */
679 NODE_DATA(nid) = careful_allocation(nid,
680 sizeof(struct pglist_data),
681 SMP_CACHE_BYTES, end_pfn);
682 NODE_DATA(nid) = __va(NODE_DATA(nid));
683 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
685 dbg("node %d\n", nid);
686 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
688 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
689 NODE_DATA(nid)->node_start_pfn = start_pfn;
690 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
692 if (NODE_DATA(nid)->node_spanned_pages == 0)
695 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
696 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
698 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
699 bootmem_paddr = (unsigned long)careful_allocation(nid,
700 bootmap_pages << PAGE_SHIFT,
702 memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
704 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
706 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
709 free_bootmem_with_active_regions(nid, end_pfn);
711 /* Mark reserved regions on this node */
712 for (i = 0; i < lmb.reserved.cnt; i++) {
713 unsigned long physbase = lmb.reserved.region[i].base;
714 unsigned long size = lmb.reserved.region[i].size;
715 unsigned long start_paddr = start_pfn << PAGE_SHIFT;
716 unsigned long end_paddr = end_pfn << PAGE_SHIFT;
718 if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
719 early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
722 if (physbase < end_paddr &&
723 (physbase+size) > start_paddr) {
725 if (physbase < start_paddr) {
726 size -= start_paddr - physbase;
727 physbase = start_paddr;
730 if (size > end_paddr - physbase)
731 size = end_paddr - physbase;
733 dbg("reserve_bootmem %lx %lx\n", physbase,
735 reserve_bootmem_node(NODE_DATA(nid), physbase,
736 size, BOOTMEM_DEFAULT);
740 sparse_memory_present_with_active_regions(nid);
744 void __init paging_init(void)
746 unsigned long max_zone_pfns[MAX_NR_ZONES];
747 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
748 max_zone_pfns[ZONE_DMA] = lmb_end_of_DRAM() >> PAGE_SHIFT;
749 free_area_init_nodes(max_zone_pfns);
752 static int __init early_numa(char *p)
757 if (strstr(p, "off"))
760 if (strstr(p, "debug"))
763 p = strstr(p, "fake=");
765 cmdline = p + strlen("fake=");
769 early_param("numa", early_numa);
771 #ifdef CONFIG_MEMORY_HOTPLUG
773 * Find the node associated with a hot added memory section. Section
774 * corresponds to a SPARSEMEM section, not an LMB. It is assumed that
775 * sections are fully contained within a single LMB.
777 int hot_add_scn_to_nid(unsigned long scn_addr)
779 struct device_node *memory = NULL;
781 int default_nid = any_online_node(NODE_MASK_ALL);
784 if (!numa_enabled || (min_common_depth < 0))
787 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
788 unsigned long start, size;
790 const unsigned int *memcell_buf;
793 memcell_buf = of_get_property(memory, "reg", &len);
794 if (!memcell_buf || len <= 0)
798 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
800 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
801 size = read_n_cells(n_mem_size_cells, &memcell_buf);
802 nid = of_node_to_nid_single(memory);
804 /* Domains not present at boot default to 0 */
805 if (nid < 0 || !node_online(nid))
808 if ((scn_addr >= start) && (scn_addr < (start + size))) {
813 if (--ranges) /* process all ranges in cell */
816 BUG(); /* section address should be found above */
819 /* Temporary code to ensure that returned node is not empty */
822 while (NODE_DATA(nid)->node_spanned_pages == 0) {
823 node_clear(nid, nodes);
824 nid = any_online_node(nodes);
828 #endif /* CONFIG_MEMORY_HOTPLUG */