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/mmzone.h>
11 #include <linux/ctype.h>
12 #include <linux/module.h>
13 #include <linux/nodemask.h>
14 #include <linux/sched.h>
17 #include <asm/proto.h>
23 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
24 EXPORT_SYMBOL(node_data);
26 struct memnode memnode;
28 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
29 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
32 int numa_off __initdata;
33 static unsigned long __initdata nodemap_addr;
34 static unsigned long __initdata nodemap_size;
36 DEFINE_PER_CPU(int, node_number) = 0;
37 EXPORT_PER_CPU_SYMBOL(node_number);
40 * Map cpu index to node index
42 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
43 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
46 * Given a shift value, try to populate memnodemap[]
49 * 0 if memnodmap[] too small (of shift too small)
50 * -1 if node overlap or lost ram (shift too big)
52 static int __init populate_memnodemap(const struct bootnode *nodes,
53 int numnodes, int shift, int *nodeids)
55 unsigned long addr, end;
58 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
59 for (i = 0; i < numnodes; i++) {
60 addr = nodes[i].start;
64 if ((end >> shift) >= memnodemapsize)
67 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
71 memnodemap[addr >> shift] = i;
73 memnodemap[addr >> shift] = nodeids[i];
75 addr += (1UL << shift);
82 static int __init allocate_cachealigned_memnodemap(void)
86 memnodemap = memnode.embedded_map;
87 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
91 nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
92 nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT,
93 nodemap_size, L1_CACHE_BYTES);
94 if (nodemap_addr == -1UL) {
96 "NUMA: Unable to allocate Memory to Node hash map\n");
97 nodemap_addr = nodemap_size = 0;
100 memnodemap = phys_to_virt(nodemap_addr);
101 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
103 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
104 nodemap_addr, nodemap_addr + nodemap_size);
109 * The LSB of all start and end addresses in the node map is the value of the
110 * maximum possible shift.
112 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
115 int i, nodes_used = 0;
116 unsigned long start, end;
117 unsigned long bitfield = 0, memtop = 0;
119 for (i = 0; i < numnodes; i++) {
120 start = nodes[i].start;
132 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
133 memnodemapsize = (memtop >> i)+1;
137 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
142 shift = extract_lsb_from_nodes(nodes, numnodes);
143 if (allocate_cachealigned_memnodemap())
145 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
148 if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
149 printk(KERN_INFO "Your memory is not aligned you need to "
150 "rebuild your kernel with a bigger NODEMAPSIZE "
151 "shift=%d\n", shift);
157 int __meminit __early_pfn_to_nid(unsigned long pfn)
159 return phys_to_nid(pfn << PAGE_SHIFT);
162 static void * __init early_node_mem(int nodeid, unsigned long start,
163 unsigned long end, unsigned long size,
166 unsigned long mem = find_e820_area(start, end, size, align);
172 start = __pa(MAX_DMA_ADDRESS);
173 end = max_low_pfn_mapped << PAGE_SHIFT;
174 mem = find_e820_area(start, end, size, align);
178 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
184 /* Initialize bootmem allocator for a node */
186 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
188 unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size;
189 const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
190 unsigned long bootmap_start, nodedata_phys;
198 * Don't confuse VM with a node that doesn't have the
199 * minimum amount of memory:
201 if (end && (end - start) < NODE_MIN_SIZE)
204 start = roundup(start, ZONE_ALIGN);
206 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
209 start_pfn = start >> PAGE_SHIFT;
210 last_pfn = end >> PAGE_SHIFT;
212 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
214 if (node_data[nodeid] == NULL)
216 nodedata_phys = __pa(node_data[nodeid]);
217 reserve_early(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
218 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
219 nodedata_phys + pgdat_size - 1);
220 nid = phys_to_nid(nodedata_phys);
222 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
224 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
225 NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid];
226 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
227 NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
230 * Find a place for the bootmem map
231 * nodedata_phys could be on other nodes by alloc_bootmem,
232 * so need to sure bootmap_start not to be small, otherwise
233 * early_node_mem will get that with find_e820_area instead
234 * of alloc_bootmem, that could clash with reserved range
236 bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn);
237 bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE);
239 * SMP_CACHE_BYTES could be enough, but init_bootmem_node like
240 * to use that to align to PAGE_SIZE
242 bootmap = early_node_mem(nodeid, bootmap_start, end,
243 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
244 if (bootmap == NULL) {
245 free_early(nodedata_phys, nodedata_phys + pgdat_size);
246 node_data[nodeid] = NULL;
249 bootmap_start = __pa(bootmap);
250 reserve_early(bootmap_start, bootmap_start+(bootmap_pages<<PAGE_SHIFT),
253 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
254 bootmap_start >> PAGE_SHIFT,
255 start_pfn, last_pfn);
257 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
258 bootmap_start, bootmap_start + bootmap_size - 1,
260 nid = phys_to_nid(bootmap_start);
262 printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid);
264 free_bootmem_with_active_regions(nodeid, end);
266 node_set_online(nodeid);
270 * There are unfortunately some poorly designed mainboards around that
271 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
272 * mapping. To avoid this fill in the mapping for all possible CPUs,
273 * as the number of CPUs is not known yet. We round robin the existing
276 void __init numa_init_array(void)
280 rr = first_node(node_online_map);
281 for (i = 0; i < nr_cpu_ids; i++) {
282 if (early_cpu_to_node(i) != NUMA_NO_NODE)
284 numa_set_node(i, rr);
285 rr = next_node(rr, node_online_map);
286 if (rr == MAX_NUMNODES)
287 rr = first_node(node_online_map);
291 #ifdef CONFIG_NUMA_EMU
293 static struct bootnode nodes[MAX_NUMNODES] __initdata;
294 static struct bootnode physnodes[MAX_NUMNODES] __initdata;
295 static char *cmdline __initdata;
297 static int __init setup_physnodes(unsigned long start, unsigned long end,
304 #ifdef CONFIG_ACPI_NUMA
306 nr_nodes = acpi_get_nodes(physnodes);
308 #ifdef CONFIG_K8_NUMA
310 nr_nodes = k8_get_nodes(physnodes);
313 * Basic sanity checking on the physical node map: there may be errors
314 * if the SRAT or K8 incorrectly reported the topology or the mem=
315 * kernel parameter is used.
317 for (i = 0; i < nr_nodes; i++) {
318 if (physnodes[i].start == physnodes[i].end)
320 if (physnodes[i].start > end) {
321 physnodes[i].end = physnodes[i].start;
324 if (physnodes[i].end < start) {
325 physnodes[i].start = physnodes[i].end;
328 if (physnodes[i].start < start)
329 physnodes[i].start = start;
330 if (physnodes[i].end > end)
331 physnodes[i].end = end;
335 * Remove all nodes that have no memory or were truncated because of the
336 * limited address range.
338 for (i = 0; i < nr_nodes; i++) {
339 if (physnodes[i].start == physnodes[i].end)
341 physnodes[ret].start = physnodes[i].start;
342 physnodes[ret].end = physnodes[i].end;
347 * If no physical topology was detected, a single node is faked to cover
348 * the entire address space.
351 physnodes[ret].start = start;
352 physnodes[ret].end = end;
359 * Setups up nid to range from addr to addr + size. If the end
360 * boundary is greater than max_addr, then max_addr is used instead.
361 * The return value is 0 if there is additional memory left for
362 * allocation past addr and -1 otherwise. addr is adjusted to be at
363 * the end of the node.
365 static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr)
368 nodes[nid].start = *addr;
370 if (*addr >= max_addr) {
374 nodes[nid].end = *addr;
375 node_set(nid, node_possible_map);
376 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
377 nodes[nid].start, nodes[nid].end,
378 (nodes[nid].end - nodes[nid].start) >> 20);
383 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
384 * to max_addr. The return value is the number of nodes allocated.
386 static int __init split_nodes_interleave(u64 addr, u64 max_addr,
387 int nr_phys_nodes, int nr_nodes)
389 nodemask_t physnode_mask = NODE_MASK_NONE;
397 if (nr_nodes > MAX_NUMNODES) {
398 pr_info("numa=fake=%d too large, reducing to %d\n",
399 nr_nodes, MAX_NUMNODES);
400 nr_nodes = MAX_NUMNODES;
403 size = (max_addr - addr - e820_hole_size(addr, max_addr)) / nr_nodes;
405 * Calculate the number of big nodes that can be allocated as a result
406 * of consolidating the remainder.
408 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) & nr_nodes) /
411 size &= FAKE_NODE_MIN_HASH_MASK;
413 pr_err("Not enough memory for each node. "
414 "NUMA emulation disabled.\n");
418 for (i = 0; i < nr_phys_nodes; i++)
419 if (physnodes[i].start != physnodes[i].end)
420 node_set(i, physnode_mask);
423 * Continue to fill physical nodes with fake nodes until there is no
424 * memory left on any of them.
426 while (nodes_weight(physnode_mask)) {
427 for_each_node_mask(i, physnode_mask) {
428 u64 end = physnodes[i].start + size;
429 u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
432 end += FAKE_NODE_MIN_SIZE;
435 * Continue to add memory to this fake node if its
436 * non-reserved memory is less than the per-node size.
438 while (end - physnodes[i].start -
439 e820_hole_size(physnodes[i].start, end) < size) {
440 end += FAKE_NODE_MIN_SIZE;
441 if (end > physnodes[i].end) {
442 end = physnodes[i].end;
448 * If there won't be at least FAKE_NODE_MIN_SIZE of
449 * non-reserved memory in ZONE_DMA32 for the next node,
450 * this one must extend to the boundary.
452 if (end < dma32_end && dma32_end - end -
453 e820_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
457 * If there won't be enough non-reserved memory for the
458 * next node, this one must extend to the end of the
461 if (physnodes[i].end - end -
462 e820_hole_size(end, physnodes[i].end) < size)
463 end = physnodes[i].end;
466 * Avoid allocating more nodes than requested, which can
467 * happen as a result of rounding down each node's size
468 * to FAKE_NODE_MIN_SIZE.
470 if (nodes_weight(physnode_mask) + ret >= nr_nodes)
471 end = physnodes[i].end;
473 if (setup_node_range(ret++, &physnodes[i].start,
474 end - physnodes[i].start,
475 physnodes[i].end) < 0)
476 node_clear(i, physnode_mask);
483 * Splits num_nodes nodes up equally starting at node_start. The return value
484 * is the number of nodes split up and addr is adjusted to be at the end of the
485 * last node allocated.
487 static int __init split_nodes_equally(u64 *addr, u64 max_addr, int node_start,
496 if (num_nodes > MAX_NUMNODES)
497 num_nodes = MAX_NUMNODES;
498 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
501 * Calculate the number of big nodes that can be allocated as a result
502 * of consolidating the leftovers.
504 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
507 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
508 size &= FAKE_NODE_MIN_HASH_MASK;
510 printk(KERN_ERR "Not enough memory for each node. "
511 "NUMA emulation disabled.\n");
515 for (i = node_start; i < num_nodes + node_start; i++) {
516 u64 end = *addr + size;
519 end += FAKE_NODE_MIN_SIZE;
521 * The final node can have the remaining system RAM. Other
522 * nodes receive roughly the same amount of available pages.
524 if (i == num_nodes + node_start - 1)
527 while (end - *addr - e820_hole_size(*addr, end) <
529 end += FAKE_NODE_MIN_SIZE;
530 if (end > max_addr) {
535 if (setup_node_range(i, addr, end - *addr, max_addr) < 0)
538 return i - node_start + 1;
542 * Splits the remaining system RAM into chunks of size. The remaining memory is
543 * always assigned to a final node and can be asymmetric. Returns the number of
546 static int __init split_nodes_by_size(u64 *addr, u64 max_addr, int node_start,
550 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
551 while (!setup_node_range(i++, addr, size, max_addr))
553 return i - node_start;
557 * Sets up the system RAM area from start_pfn to last_pfn according to the
558 * numa=fake command-line option.
560 static int __init numa_emulation(unsigned long start_pfn,
561 unsigned long last_pfn, int acpi, int k8)
563 u64 size, addr = start_pfn << PAGE_SHIFT;
564 u64 max_addr = last_pfn << PAGE_SHIFT;
565 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
568 num_phys_nodes = setup_physnodes(addr, max_addr, acpi, k8);
570 * If the numa=fake command-line is just a single number N, split the
571 * system RAM into N fake nodes.
573 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
574 long n = simple_strtol(cmdline, NULL, 0);
576 num_nodes = split_nodes_interleave(addr, max_addr,
583 /* Parse the command line. */
584 for (coeff_flag = 0; ; cmdline++) {
585 if (*cmdline && isdigit(*cmdline)) {
586 num = num * 10 + *cmdline - '0';
589 if (*cmdline == '*') {
594 if (!*cmdline || *cmdline == ',') {
598 * Round down to the nearest FAKE_NODE_MIN_SIZE.
599 * Command-line coefficients are in megabytes.
601 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
603 for (i = 0; i < coeff; i++, num_nodes++)
604 if (setup_node_range(num_nodes, &addr,
617 /* Fill remainder of system RAM, if appropriate. */
618 if (addr < max_addr) {
619 if (coeff_flag && coeff < 0) {
620 /* Split remaining nodes into num-sized chunks */
621 num_nodes += split_nodes_by_size(&addr, max_addr,
625 switch (*(cmdline - 1)) {
627 /* Split remaining nodes into coeff chunks */
630 num_nodes += split_nodes_equally(&addr, max_addr,
634 /* Do not allocate remaining system RAM */
637 /* Give one final node */
638 setup_node_range(num_nodes, &addr, max_addr - addr,
644 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
645 if (memnode_shift < 0) {
647 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
653 * We need to vacate all active ranges that may have been registered for
654 * the e820 memory map.
656 remove_all_active_ranges();
657 for_each_node_mask(i, node_possible_map) {
658 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
659 nodes[i].end >> PAGE_SHIFT);
660 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
662 acpi_fake_nodes(nodes, num_nodes);
666 #endif /* CONFIG_NUMA_EMU */
668 void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn,
673 nodes_clear(node_possible_map);
674 nodes_clear(node_online_map);
676 #ifdef CONFIG_NUMA_EMU
677 if (cmdline && !numa_emulation(start_pfn, last_pfn, acpi, k8))
679 nodes_clear(node_possible_map);
680 nodes_clear(node_online_map);
683 #ifdef CONFIG_ACPI_NUMA
684 if (!numa_off && acpi && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
685 last_pfn << PAGE_SHIFT))
687 nodes_clear(node_possible_map);
688 nodes_clear(node_online_map);
691 #ifdef CONFIG_K8_NUMA
692 if (!numa_off && k8 && !k8_scan_nodes())
694 nodes_clear(node_possible_map);
695 nodes_clear(node_online_map);
697 printk(KERN_INFO "%s\n",
698 numa_off ? "NUMA turned off" : "No NUMA configuration found");
700 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
701 start_pfn << PAGE_SHIFT,
702 last_pfn << PAGE_SHIFT);
703 /* setup dummy node covering all memory */
705 memnodemap = memnode.embedded_map;
708 node_set(0, node_possible_map);
709 for (i = 0; i < nr_cpu_ids; i++)
711 e820_register_active_regions(0, start_pfn, last_pfn);
712 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
715 unsigned long __init numa_free_all_bootmem(void)
717 unsigned long pages = 0;
720 for_each_online_node(i)
721 pages += free_all_bootmem_node(NODE_DATA(i));
726 static __init int numa_setup(char *opt)
730 if (!strncmp(opt, "off", 3))
732 #ifdef CONFIG_NUMA_EMU
733 if (!strncmp(opt, "fake=", 5))
736 #ifdef CONFIG_ACPI_NUMA
737 if (!strncmp(opt, "noacpi", 6))
742 early_param("numa", numa_setup);
746 static __init int find_near_online_node(int node)
749 int min_val = INT_MAX;
752 for_each_online_node(n) {
753 val = node_distance(node, n);
765 * Setup early cpu_to_node.
767 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
768 * and apicid_to_node[] tables have valid entries for a CPU.
769 * This means we skip cpu_to_node[] initialisation for NUMA
770 * emulation and faking node case (when running a kernel compiled
771 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
772 * is already initialized in a round robin manner at numa_init_array,
773 * prior to this call, and this initialization is good enough
774 * for the fake NUMA cases.
776 * Called before the per_cpu areas are setup.
778 void __init init_cpu_to_node(void)
781 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
783 BUG_ON(cpu_to_apicid == NULL);
785 for_each_possible_cpu(cpu) {
787 u16 apicid = cpu_to_apicid[cpu];
789 if (apicid == BAD_APICID)
791 node = apicid_to_node[apicid];
792 if (node == NUMA_NO_NODE)
794 if (!node_online(node))
795 node = find_near_online_node(node);
796 numa_set_node(cpu, node);
802 void __cpuinit numa_set_node(int cpu, int node)
804 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
806 /* early setting, no percpu area yet */
807 if (cpu_to_node_map) {
808 cpu_to_node_map[cpu] = node;
812 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
813 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
814 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
819 per_cpu(x86_cpu_to_node_map, cpu) = node;
821 if (node != NUMA_NO_NODE)
822 per_cpu(node_number, cpu) = node;
825 void __cpuinit numa_clear_node(int cpu)
827 numa_set_node(cpu, NUMA_NO_NODE);
830 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
832 void __cpuinit numa_add_cpu(int cpu)
834 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
837 void __cpuinit numa_remove_cpu(int cpu)
839 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
842 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
845 * --------- debug versions of the numa functions ---------
847 static void __cpuinit numa_set_cpumask(int cpu, int enable)
849 int node = early_cpu_to_node(cpu);
850 struct cpumask *mask;
853 mask = node_to_cpumask_map[node];
855 printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node);
861 cpumask_set_cpu(cpu, mask);
863 cpumask_clear_cpu(cpu, mask);
865 cpulist_scnprintf(buf, sizeof(buf), mask);
866 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
867 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
870 void __cpuinit numa_add_cpu(int cpu)
872 numa_set_cpumask(cpu, 1);
875 void __cpuinit numa_remove_cpu(int cpu)
877 numa_set_cpumask(cpu, 0);
880 int cpu_to_node(int cpu)
882 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
884 "cpu_to_node(%d): usage too early!\n", cpu);
886 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
888 return per_cpu(x86_cpu_to_node_map, cpu);
890 EXPORT_SYMBOL(cpu_to_node);
893 * Same function as cpu_to_node() but used if called before the
894 * per_cpu areas are setup.
896 int early_cpu_to_node(int cpu)
898 if (early_per_cpu_ptr(x86_cpu_to_node_map))
899 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
901 if (!cpu_possible(cpu)) {
903 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
907 return per_cpu(x86_cpu_to_node_map, cpu);
911 * --------- end of debug versions of the numa functions ---------
914 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */