2 * Generic VM initialization for x86-64 NUMA setups.
3 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 #include <linux/kernel.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/memblock.h>
11 #include <linux/mmzone.h>
12 #include <linux/ctype.h>
13 #include <linux/module.h>
14 #include <linux/nodemask.h>
15 #include <linux/sched.h>
18 #include <asm/proto.h>
22 #include <asm/amd_nb.h>
24 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
25 EXPORT_SYMBOL(node_data);
27 struct memnode memnode;
29 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
30 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
33 int numa_off __initdata;
34 static unsigned long __initdata nodemap_addr;
35 static unsigned long __initdata nodemap_size;
38 * Map cpu index to node index
40 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
41 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
44 * Given a shift value, try to populate memnodemap[]
47 * 0 if memnodmap[] too small (of shift too small)
48 * -1 if node overlap or lost ram (shift too big)
50 static int __init populate_memnodemap(const struct bootnode *nodes,
51 int numnodes, int shift, int *nodeids)
53 unsigned long addr, end;
56 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
57 for (i = 0; i < numnodes; i++) {
58 addr = nodes[i].start;
62 if ((end >> shift) >= memnodemapsize)
65 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
69 memnodemap[addr >> shift] = i;
71 memnodemap[addr >> shift] = nodeids[i];
73 addr += (1UL << shift);
80 static int __init allocate_cachealigned_memnodemap(void)
84 memnodemap = memnode.embedded_map;
85 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
89 nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
90 nodemap_addr = memblock_find_in_range(addr, max_pfn<<PAGE_SHIFT,
91 nodemap_size, L1_CACHE_BYTES);
92 if (nodemap_addr == MEMBLOCK_ERROR) {
94 "NUMA: Unable to allocate Memory to Node hash map\n");
95 nodemap_addr = nodemap_size = 0;
98 memnodemap = phys_to_virt(nodemap_addr);
99 memblock_x86_reserve_range(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
101 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
102 nodemap_addr, nodemap_addr + nodemap_size);
107 * The LSB of all start and end addresses in the node map is the value of the
108 * maximum possible shift.
110 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
113 int i, nodes_used = 0;
114 unsigned long start, end;
115 unsigned long bitfield = 0, memtop = 0;
117 for (i = 0; i < numnodes; i++) {
118 start = nodes[i].start;
130 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
131 memnodemapsize = (memtop >> i)+1;
135 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
140 shift = extract_lsb_from_nodes(nodes, numnodes);
141 if (allocate_cachealigned_memnodemap())
143 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
146 if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
147 printk(KERN_INFO "Your memory is not aligned you need to "
148 "rebuild your kernel with a bigger NODEMAPSIZE "
149 "shift=%d\n", shift);
155 int __meminit __early_pfn_to_nid(unsigned long pfn)
157 return phys_to_nid(pfn << PAGE_SHIFT);
160 static void * __init early_node_mem(int nodeid, unsigned long start,
161 unsigned long end, unsigned long size,
167 * put it on high as possible
168 * something will go with NODE_DATA
170 if (start < (MAX_DMA_PFN<<PAGE_SHIFT))
171 start = MAX_DMA_PFN<<PAGE_SHIFT;
172 if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) &&
173 end > (MAX_DMA32_PFN<<PAGE_SHIFT))
174 start = MAX_DMA32_PFN<<PAGE_SHIFT;
175 mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
176 if (mem != MEMBLOCK_ERROR)
179 /* extend the search scope */
180 end = max_pfn_mapped << PAGE_SHIFT;
181 if (end > (MAX_DMA32_PFN<<PAGE_SHIFT))
182 start = MAX_DMA32_PFN<<PAGE_SHIFT;
184 start = MAX_DMA_PFN<<PAGE_SHIFT;
185 mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
186 if (mem != MEMBLOCK_ERROR)
189 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
195 /* Initialize bootmem allocator for a node */
197 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
199 unsigned long start_pfn, last_pfn, nodedata_phys;
200 const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
207 * Don't confuse VM with a node that doesn't have the
208 * minimum amount of memory:
210 if (end && (end - start) < NODE_MIN_SIZE)
213 start = roundup(start, ZONE_ALIGN);
215 printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid,
218 start_pfn = start >> PAGE_SHIFT;
219 last_pfn = end >> PAGE_SHIFT;
221 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
223 if (node_data[nodeid] == NULL)
225 nodedata_phys = __pa(node_data[nodeid]);
226 memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
227 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
228 nodedata_phys + pgdat_size - 1);
229 nid = phys_to_nid(nodedata_phys);
231 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
233 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
234 NODE_DATA(nodeid)->node_id = nodeid;
235 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
236 NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
238 node_set_online(nodeid);
242 * There are unfortunately some poorly designed mainboards around that
243 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
244 * mapping. To avoid this fill in the mapping for all possible CPUs,
245 * as the number of CPUs is not known yet. We round robin the existing
248 void __init numa_init_array(void)
252 rr = first_node(node_online_map);
253 for (i = 0; i < nr_cpu_ids; i++) {
254 if (early_cpu_to_node(i) != NUMA_NO_NODE)
256 numa_set_node(i, rr);
257 rr = next_node(rr, node_online_map);
258 if (rr == MAX_NUMNODES)
259 rr = first_node(node_online_map);
263 #ifdef CONFIG_NUMA_EMU
265 static struct bootnode nodes[MAX_NUMNODES] __initdata;
266 static struct bootnode physnodes[MAX_NUMNODES] __initdata;
267 static char *cmdline __initdata;
269 static int __init setup_physnodes(unsigned long start, unsigned long end,
276 #ifdef CONFIG_ACPI_NUMA
278 nr_nodes = acpi_get_nodes(physnodes);
280 #ifdef CONFIG_K8_NUMA
282 nr_nodes = k8_get_nodes(physnodes);
285 * Basic sanity checking on the physical node map: there may be errors
286 * if the SRAT or K8 incorrectly reported the topology or the mem=
287 * kernel parameter is used.
289 for (i = 0; i < nr_nodes; i++) {
290 if (physnodes[i].start == physnodes[i].end)
292 if (physnodes[i].start > end) {
293 physnodes[i].end = physnodes[i].start;
296 if (physnodes[i].end < start) {
297 physnodes[i].start = physnodes[i].end;
300 if (physnodes[i].start < start)
301 physnodes[i].start = start;
302 if (physnodes[i].end > end)
303 physnodes[i].end = end;
307 * Remove all nodes that have no memory or were truncated because of the
308 * limited address range.
310 for (i = 0; i < nr_nodes; i++) {
311 if (physnodes[i].start == physnodes[i].end)
313 physnodes[ret].start = physnodes[i].start;
314 physnodes[ret].end = physnodes[i].end;
319 * If no physical topology was detected, a single node is faked to cover
320 * the entire address space.
323 physnodes[ret].start = start;
324 physnodes[ret].end = end;
331 * Setups up nid to range from addr to addr + size. If the end
332 * boundary is greater than max_addr, then max_addr is used instead.
333 * The return value is 0 if there is additional memory left for
334 * allocation past addr and -1 otherwise. addr is adjusted to be at
335 * the end of the node.
337 static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr)
340 nodes[nid].start = *addr;
342 if (*addr >= max_addr) {
346 nodes[nid].end = *addr;
347 node_set(nid, node_possible_map);
348 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
349 nodes[nid].start, nodes[nid].end,
350 (nodes[nid].end - nodes[nid].start) >> 20);
355 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
356 * to max_addr. The return value is the number of nodes allocated.
358 static int __init split_nodes_interleave(u64 addr, u64 max_addr,
359 int nr_phys_nodes, int nr_nodes)
361 nodemask_t physnode_mask = NODE_MASK_NONE;
369 if (nr_nodes > MAX_NUMNODES) {
370 pr_info("numa=fake=%d too large, reducing to %d\n",
371 nr_nodes, MAX_NUMNODES);
372 nr_nodes = MAX_NUMNODES;
375 size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / nr_nodes;
377 * Calculate the number of big nodes that can be allocated as a result
378 * of consolidating the remainder.
380 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
383 size &= FAKE_NODE_MIN_HASH_MASK;
385 pr_err("Not enough memory for each node. "
386 "NUMA emulation disabled.\n");
390 for (i = 0; i < nr_phys_nodes; i++)
391 if (physnodes[i].start != physnodes[i].end)
392 node_set(i, physnode_mask);
395 * Continue to fill physical nodes with fake nodes until there is no
396 * memory left on any of them.
398 while (nodes_weight(physnode_mask)) {
399 for_each_node_mask(i, physnode_mask) {
400 u64 end = physnodes[i].start + size;
401 u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
404 end += FAKE_NODE_MIN_SIZE;
407 * Continue to add memory to this fake node if its
408 * non-reserved memory is less than the per-node size.
410 while (end - physnodes[i].start -
411 memblock_x86_hole_size(physnodes[i].start, end) < size) {
412 end += FAKE_NODE_MIN_SIZE;
413 if (end > physnodes[i].end) {
414 end = physnodes[i].end;
420 * If there won't be at least FAKE_NODE_MIN_SIZE of
421 * non-reserved memory in ZONE_DMA32 for the next node,
422 * this one must extend to the boundary.
424 if (end < dma32_end && dma32_end - end -
425 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
429 * If there won't be enough non-reserved memory for the
430 * next node, this one must extend to the end of the
433 if (physnodes[i].end - end -
434 memblock_x86_hole_size(end, physnodes[i].end) < size)
435 end = physnodes[i].end;
438 * Avoid allocating more nodes than requested, which can
439 * happen as a result of rounding down each node's size
440 * to FAKE_NODE_MIN_SIZE.
442 if (nodes_weight(physnode_mask) + ret >= nr_nodes)
443 end = physnodes[i].end;
445 if (setup_node_range(ret++, &physnodes[i].start,
446 end - physnodes[i].start,
447 physnodes[i].end) < 0)
448 node_clear(i, physnode_mask);
455 * Returns the end address of a node so that there is at least `size' amount of
456 * non-reserved memory or `max_addr' is reached.
458 static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
460 u64 end = start + size;
462 while (end - start - memblock_x86_hole_size(start, end) < size) {
463 end += FAKE_NODE_MIN_SIZE;
464 if (end > max_addr) {
473 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
474 * `addr' to `max_addr'. The return value is the number of nodes allocated.
476 static int __init split_nodes_size_interleave(u64 addr, u64 max_addr, u64 size)
478 nodemask_t physnode_mask = NODE_MASK_NONE;
486 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
487 * increased accordingly if the requested size is too small. This
488 * creates a uniform distribution of node sizes across the entire
489 * machine (but not necessarily over physical nodes).
491 min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) /
493 min_size = max(min_size, FAKE_NODE_MIN_SIZE);
494 if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
495 min_size = (min_size + FAKE_NODE_MIN_SIZE) &
496 FAKE_NODE_MIN_HASH_MASK;
497 if (size < min_size) {
498 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
499 size >> 20, min_size >> 20);
502 size &= FAKE_NODE_MIN_HASH_MASK;
504 for (i = 0; i < MAX_NUMNODES; i++)
505 if (physnodes[i].start != physnodes[i].end)
506 node_set(i, physnode_mask);
508 * Fill physical nodes with fake nodes of size until there is no memory
509 * left on any of them.
511 while (nodes_weight(physnode_mask)) {
512 for_each_node_mask(i, physnode_mask) {
513 u64 dma32_end = MAX_DMA32_PFN << PAGE_SHIFT;
516 end = find_end_of_node(physnodes[i].start,
517 physnodes[i].end, size);
519 * If there won't be at least FAKE_NODE_MIN_SIZE of
520 * non-reserved memory in ZONE_DMA32 for the next node,
521 * this one must extend to the boundary.
523 if (end < dma32_end && dma32_end - end -
524 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
528 * If there won't be enough non-reserved memory for the
529 * next node, this one must extend to the end of the
532 if (physnodes[i].end - end -
533 memblock_x86_hole_size(end, physnodes[i].end) < size)
534 end = physnodes[i].end;
537 * Setup the fake node that will be allocated as bootmem
538 * later. If setup_node_range() returns non-zero, there
539 * is no more memory available on this physical node.
541 if (setup_node_range(ret++, &physnodes[i].start,
542 end - physnodes[i].start,
543 physnodes[i].end) < 0)
544 node_clear(i, physnode_mask);
551 * Sets up the system RAM area from start_pfn to last_pfn according to the
552 * numa=fake command-line option.
554 static int __init numa_emulation(unsigned long start_pfn,
555 unsigned long last_pfn, int acpi, int k8)
557 u64 addr = start_pfn << PAGE_SHIFT;
558 u64 max_addr = last_pfn << PAGE_SHIFT;
563 num_phys_nodes = setup_physnodes(addr, max_addr, acpi, k8);
565 * If the numa=fake command-line contains a 'M' or 'G', it represents
566 * the fixed node size. Otherwise, if it is just a single number N,
567 * split the system RAM into N fake nodes.
569 if (strchr(cmdline, 'M') || strchr(cmdline, 'G')) {
572 size = memparse(cmdline, &cmdline);
573 num_nodes = split_nodes_size_interleave(addr, max_addr, size);
577 n = simple_strtoul(cmdline, NULL, 0);
578 num_nodes = split_nodes_interleave(addr, max_addr, num_phys_nodes, n);
583 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
584 if (memnode_shift < 0) {
586 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
592 * We need to vacate all active ranges that may have been registered for
593 * the e820 memory map.
595 remove_all_active_ranges();
596 for_each_node_mask(i, node_possible_map) {
597 memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
598 nodes[i].end >> PAGE_SHIFT);
599 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
601 acpi_fake_nodes(nodes, num_nodes);
605 #endif /* CONFIG_NUMA_EMU */
607 void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn,
612 nodes_clear(node_possible_map);
613 nodes_clear(node_online_map);
615 #ifdef CONFIG_NUMA_EMU
616 if (cmdline && !numa_emulation(start_pfn, last_pfn, acpi, k8))
618 nodes_clear(node_possible_map);
619 nodes_clear(node_online_map);
622 #ifdef CONFIG_ACPI_NUMA
623 if (!numa_off && acpi && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
624 last_pfn << PAGE_SHIFT))
626 nodes_clear(node_possible_map);
627 nodes_clear(node_online_map);
630 #ifdef CONFIG_K8_NUMA
631 if (!numa_off && k8 && !k8_scan_nodes())
633 nodes_clear(node_possible_map);
634 nodes_clear(node_online_map);
636 printk(KERN_INFO "%s\n",
637 numa_off ? "NUMA turned off" : "No NUMA configuration found");
639 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
640 start_pfn << PAGE_SHIFT,
641 last_pfn << PAGE_SHIFT);
642 /* setup dummy node covering all memory */
644 memnodemap = memnode.embedded_map;
647 node_set(0, node_possible_map);
648 for (i = 0; i < nr_cpu_ids; i++)
650 memblock_x86_register_active_regions(0, start_pfn, last_pfn);
651 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
654 unsigned long __init numa_free_all_bootmem(void)
656 unsigned long pages = 0;
659 for_each_online_node(i)
660 pages += free_all_bootmem_node(NODE_DATA(i));
662 pages += free_all_memory_core_early(MAX_NUMNODES);
667 static __init int numa_setup(char *opt)
671 if (!strncmp(opt, "off", 3))
673 #ifdef CONFIG_NUMA_EMU
674 if (!strncmp(opt, "fake=", 5))
677 #ifdef CONFIG_ACPI_NUMA
678 if (!strncmp(opt, "noacpi", 6))
683 early_param("numa", numa_setup);
687 static __init int find_near_online_node(int node)
690 int min_val = INT_MAX;
693 for_each_online_node(n) {
694 val = node_distance(node, n);
706 * Setup early cpu_to_node.
708 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
709 * and apicid_to_node[] tables have valid entries for a CPU.
710 * This means we skip cpu_to_node[] initialisation for NUMA
711 * emulation and faking node case (when running a kernel compiled
712 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
713 * is already initialized in a round robin manner at numa_init_array,
714 * prior to this call, and this initialization is good enough
715 * for the fake NUMA cases.
717 * Called before the per_cpu areas are setup.
719 void __init init_cpu_to_node(void)
722 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
724 BUG_ON(cpu_to_apicid == NULL);
726 for_each_possible_cpu(cpu) {
728 u16 apicid = cpu_to_apicid[cpu];
730 if (apicid == BAD_APICID)
732 node = apicid_to_node[apicid];
733 if (node == NUMA_NO_NODE)
735 if (!node_online(node))
736 node = find_near_online_node(node);
737 numa_set_node(cpu, node);
743 void __cpuinit numa_set_node(int cpu, int node)
745 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
747 /* early setting, no percpu area yet */
748 if (cpu_to_node_map) {
749 cpu_to_node_map[cpu] = node;
753 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
754 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
755 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
760 per_cpu(x86_cpu_to_node_map, cpu) = node;
762 if (node != NUMA_NO_NODE)
763 set_cpu_numa_node(cpu, node);
766 void __cpuinit numa_clear_node(int cpu)
768 numa_set_node(cpu, NUMA_NO_NODE);
771 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
773 void __cpuinit numa_add_cpu(int cpu)
775 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
778 void __cpuinit numa_remove_cpu(int cpu)
780 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
783 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
786 * --------- debug versions of the numa functions ---------
788 static void __cpuinit numa_set_cpumask(int cpu, int enable)
790 int node = early_cpu_to_node(cpu);
791 struct cpumask *mask;
794 mask = node_to_cpumask_map[node];
796 printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node);
802 cpumask_set_cpu(cpu, mask);
804 cpumask_clear_cpu(cpu, mask);
806 cpulist_scnprintf(buf, sizeof(buf), mask);
807 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
808 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
811 void __cpuinit numa_add_cpu(int cpu)
813 numa_set_cpumask(cpu, 1);
816 void __cpuinit numa_remove_cpu(int cpu)
818 numa_set_cpumask(cpu, 0);
821 int __cpu_to_node(int cpu)
823 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
825 "cpu_to_node(%d): usage too early!\n", cpu);
827 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
829 return per_cpu(x86_cpu_to_node_map, cpu);
831 EXPORT_SYMBOL(__cpu_to_node);
834 * Same function as cpu_to_node() but used if called before the
835 * per_cpu areas are setup.
837 int early_cpu_to_node(int cpu)
839 if (early_per_cpu_ptr(x86_cpu_to_node_map))
840 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
842 if (!cpu_possible(cpu)) {
844 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
848 return per_cpu(x86_cpu_to_node_map, cpu);
852 * --------- end of debug versions of the numa functions ---------
855 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */