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>
27 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
28 EXPORT_SYMBOL(node_data);
30 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
32 struct memnode memnode;
35 int x86_cpu_to_node_map_init[NR_CPUS] = {
36 [0 ... NR_CPUS-1] = NUMA_NO_NODE
38 void *x86_cpu_to_node_map_early_ptr;
39 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr);
41 DEFINE_PER_CPU(int, x86_cpu_to_node_map) = NUMA_NO_NODE;
42 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map);
44 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
45 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
48 cpumask_t node_to_cpumask_map[MAX_NUMNODES] __read_mostly;
49 EXPORT_SYMBOL(node_to_cpumask_map);
51 int numa_off __initdata;
52 unsigned long __initdata nodemap_addr;
53 unsigned long __initdata nodemap_size;
56 * Given a shift value, try to populate memnodemap[]
59 * 0 if memnodmap[] too small (of shift too small)
60 * -1 if node overlap or lost ram (shift too big)
62 static int __init populate_memnodemap(const struct bootnode *nodes,
63 int numnodes, int shift, int *nodeids)
65 unsigned long addr, end;
68 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
69 for (i = 0; i < numnodes; i++) {
70 addr = nodes[i].start;
74 if ((end >> shift) >= memnodemapsize)
77 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
81 memnodemap[addr >> shift] = i;
83 memnodemap[addr >> shift] = nodeids[i];
85 addr += (1UL << shift);
92 static int __init allocate_cachealigned_memnodemap(void)
96 memnodemap = memnode.embedded_map;
97 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
101 nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
102 nodemap_addr = find_e820_area(addr, end_pfn<<PAGE_SHIFT,
103 nodemap_size, L1_CACHE_BYTES);
104 if (nodemap_addr == -1UL) {
106 "NUMA: Unable to allocate Memory to Node hash map\n");
107 nodemap_addr = nodemap_size = 0;
110 memnodemap = phys_to_virt(nodemap_addr);
111 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
113 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
114 nodemap_addr, nodemap_addr + nodemap_size);
119 * The LSB of all start and end addresses in the node map is the value of the
120 * maximum possible shift.
122 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
125 int i, nodes_used = 0;
126 unsigned long start, end;
127 unsigned long bitfield = 0, memtop = 0;
129 for (i = 0; i < numnodes; i++) {
130 start = nodes[i].start;
142 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
143 memnodemapsize = (memtop >> i)+1;
147 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
152 shift = extract_lsb_from_nodes(nodes, numnodes);
153 if (allocate_cachealigned_memnodemap())
155 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
158 if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
159 printk(KERN_INFO "Your memory is not aligned you need to "
160 "rebuild your kernel with a bigger NODEMAPSIZE "
161 "shift=%d\n", shift);
167 int early_pfn_to_nid(unsigned long pfn)
169 return phys_to_nid(pfn << PAGE_SHIFT);
172 static void * __init early_node_mem(int nodeid, unsigned long start,
173 unsigned long end, unsigned long size,
176 unsigned long mem = find_e820_area(start, end, size, align);
182 ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
184 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
191 /* Initialize bootmem allocator for a node */
192 void __init setup_node_bootmem(int nodeid, unsigned long start,
195 unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
196 unsigned long bootmap_start, nodedata_phys;
198 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
201 start = round_up(start, ZONE_ALIGN);
203 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
206 start_pfn = start >> PAGE_SHIFT;
207 end_pfn = end >> PAGE_SHIFT;
209 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
211 if (node_data[nodeid] == NULL)
213 nodedata_phys = __pa(node_data[nodeid]);
214 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
215 nodedata_phys + pgdat_size - 1);
217 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
218 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
219 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
220 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
223 * Find a place for the bootmem map
224 * nodedata_phys could be on other nodes by alloc_bootmem,
225 * so need to sure bootmap_start not to be small, otherwise
226 * early_node_mem will get that with find_e820_area instead
227 * of alloc_bootmem, that could clash with reserved range
229 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
230 nid = phys_to_nid(nodedata_phys);
232 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
234 bootmap_start = round_up(start, PAGE_SIZE);
236 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
237 * to use that to align to PAGE_SIZE
239 bootmap = early_node_mem(nodeid, bootmap_start, end,
240 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
241 if (bootmap == NULL) {
242 if (nodedata_phys < start || nodedata_phys >= end)
243 free_bootmem(nodedata_phys, pgdat_size);
244 node_data[nodeid] = NULL;
247 bootmap_start = __pa(bootmap);
249 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
250 bootmap_start >> PAGE_SHIFT,
253 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
254 bootmap_start, bootmap_start + bootmap_size - 1,
257 free_bootmem_with_active_regions(nodeid, end);
260 * convert early reserve to bootmem reserve earlier
261 * otherwise early_node_mem could use early reserved mem
264 early_res_to_bootmem(start, end);
267 * in some case early_node_mem could use alloc_bootmem
268 * to get range on other node, don't reserve that again
271 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
273 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys,
274 pgdat_size, BOOTMEM_DEFAULT);
275 nid = phys_to_nid(bootmap_start);
277 printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid);
279 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
280 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
282 #ifdef CONFIG_ACPI_NUMA
283 srat_reserve_add_area(nodeid);
285 node_set_online(nodeid);
289 * There are unfortunately some poorly designed mainboards around that
290 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
291 * mapping. To avoid this fill in the mapping for all possible CPUs,
292 * as the number of CPUs is not known yet. We round robin the existing
295 void __init numa_init_array(void)
299 rr = first_node(node_online_map);
300 for (i = 0; i < NR_CPUS; i++) {
301 if (early_cpu_to_node(i) != NUMA_NO_NODE)
303 numa_set_node(i, rr);
304 rr = next_node(rr, node_online_map);
305 if (rr == MAX_NUMNODES)
306 rr = first_node(node_online_map);
310 #ifdef CONFIG_NUMA_EMU
312 char *cmdline __initdata;
315 * Setups up nid to range from addr to addr + size. If the end
316 * boundary is greater than max_addr, then max_addr is used instead.
317 * The return value is 0 if there is additional memory left for
318 * allocation past addr and -1 otherwise. addr is adjusted to be at
319 * the end of the node.
321 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
322 u64 size, u64 max_addr)
326 nodes[nid].start = *addr;
328 if (*addr >= max_addr) {
332 nodes[nid].end = *addr;
333 node_set(nid, node_possible_map);
334 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
335 nodes[nid].start, nodes[nid].end,
336 (nodes[nid].end - nodes[nid].start) >> 20);
341 * Splits num_nodes nodes up equally starting at node_start. The return value
342 * is the number of nodes split up and addr is adjusted to be at the end of the
343 * last node allocated.
345 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
346 u64 max_addr, int node_start,
355 if (num_nodes > MAX_NUMNODES)
356 num_nodes = MAX_NUMNODES;
357 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
360 * Calculate the number of big nodes that can be allocated as a result
361 * of consolidating the leftovers.
363 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
366 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
367 size &= FAKE_NODE_MIN_HASH_MASK;
369 printk(KERN_ERR "Not enough memory for each node. "
370 "NUMA emulation disabled.\n");
374 for (i = node_start; i < num_nodes + node_start; i++) {
375 u64 end = *addr + size;
378 end += FAKE_NODE_MIN_SIZE;
380 * The final node can have the remaining system RAM. Other
381 * nodes receive roughly the same amount of available pages.
383 if (i == num_nodes + node_start - 1)
386 while (end - *addr - e820_hole_size(*addr, end) <
388 end += FAKE_NODE_MIN_SIZE;
389 if (end > max_addr) {
394 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
397 return i - node_start + 1;
401 * Splits the remaining system RAM into chunks of size. The remaining memory is
402 * always assigned to a final node and can be asymmetric. Returns the number of
405 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
406 u64 max_addr, int node_start, u64 size)
409 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
410 while (!setup_node_range(i++, nodes, addr, size, max_addr))
412 return i - node_start;
416 * Sets up the system RAM area from start_pfn to end_pfn according to the
417 * numa=fake command-line option.
419 static struct bootnode nodes[MAX_NUMNODES] __initdata;
421 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
423 u64 size, addr = start_pfn << PAGE_SHIFT;
424 u64 max_addr = end_pfn << PAGE_SHIFT;
425 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
427 memset(&nodes, 0, sizeof(nodes));
429 * If the numa=fake command-line is just a single number N, split the
430 * system RAM into N fake nodes.
432 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
433 long n = simple_strtol(cmdline, NULL, 0);
435 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
441 /* Parse the command line. */
442 for (coeff_flag = 0; ; cmdline++) {
443 if (*cmdline && isdigit(*cmdline)) {
444 num = num * 10 + *cmdline - '0';
447 if (*cmdline == '*') {
452 if (!*cmdline || *cmdline == ',') {
456 * Round down to the nearest FAKE_NODE_MIN_SIZE.
457 * Command-line coefficients are in megabytes.
459 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
461 for (i = 0; i < coeff; i++, num_nodes++)
462 if (setup_node_range(num_nodes, nodes,
463 &addr, size, max_addr) < 0)
475 /* Fill remainder of system RAM, if appropriate. */
476 if (addr < max_addr) {
477 if (coeff_flag && coeff < 0) {
478 /* Split remaining nodes into num-sized chunks */
479 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
483 switch (*(cmdline - 1)) {
485 /* Split remaining nodes into coeff chunks */
488 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
492 /* Do not allocate remaining system RAM */
495 /* Give one final node */
496 setup_node_range(num_nodes, nodes, &addr,
497 max_addr - addr, max_addr);
502 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
503 if (memnode_shift < 0) {
505 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
511 * We need to vacate all active ranges that may have been registered by
512 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
513 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
515 remove_all_active_ranges();
516 #ifdef CONFIG_ACPI_NUMA
519 for_each_node_mask(i, node_possible_map) {
520 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
521 nodes[i].end >> PAGE_SHIFT);
522 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
524 acpi_fake_nodes(nodes, num_nodes);
528 #endif /* CONFIG_NUMA_EMU */
530 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
534 nodes_clear(node_possible_map);
535 nodes_clear(node_online_map);
537 #ifdef CONFIG_NUMA_EMU
538 if (cmdline && !numa_emulation(start_pfn, end_pfn))
540 nodes_clear(node_possible_map);
541 nodes_clear(node_online_map);
544 #ifdef CONFIG_ACPI_NUMA
545 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
546 end_pfn << PAGE_SHIFT))
548 nodes_clear(node_possible_map);
549 nodes_clear(node_online_map);
552 #ifdef CONFIG_K8_NUMA
553 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
554 end_pfn<<PAGE_SHIFT))
556 nodes_clear(node_possible_map);
557 nodes_clear(node_online_map);
559 printk(KERN_INFO "%s\n",
560 numa_off ? "NUMA turned off" : "No NUMA configuration found");
562 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
563 start_pfn << PAGE_SHIFT,
564 end_pfn << PAGE_SHIFT);
565 /* setup dummy node covering all memory */
567 memnodemap = memnode.embedded_map;
570 node_set(0, node_possible_map);
571 for (i = 0; i < NR_CPUS; i++)
573 /* cpumask_of_cpu() may not be available during early startup */
574 memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
575 cpu_set(0, node_to_cpumask_map[0]);
576 e820_register_active_regions(0, start_pfn, end_pfn);
577 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
580 __cpuinit void numa_add_cpu(int cpu)
583 (unsigned long *)&node_to_cpumask_map[early_cpu_to_node(cpu)]);
586 void __cpuinit numa_set_node(int cpu, int node)
588 int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
591 cpu_to_node_map[cpu] = node;
592 else if(per_cpu_offset(cpu))
593 per_cpu(x86_cpu_to_node_map, cpu) = node;
595 Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
598 unsigned long __init numa_free_all_bootmem(void)
600 unsigned long pages = 0;
603 for_each_online_node(i)
604 pages += free_all_bootmem_node(NODE_DATA(i));
609 void __init paging_init(void)
611 unsigned long max_zone_pfns[MAX_NR_ZONES];
613 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
614 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
615 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
616 max_zone_pfns[ZONE_NORMAL] = end_pfn;
618 sparse_memory_present_with_active_regions(MAX_NUMNODES);
621 free_area_init_nodes(max_zone_pfns);
624 static __init int numa_setup(char *opt)
628 if (!strncmp(opt, "off", 3))
630 #ifdef CONFIG_NUMA_EMU
631 if (!strncmp(opt, "fake=", 5))
634 #ifdef CONFIG_ACPI_NUMA
635 if (!strncmp(opt, "noacpi", 6))
637 if (!strncmp(opt, "hotadd=", 7))
638 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
642 early_param("numa", numa_setup);
645 * Setup early cpu_to_node.
647 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
648 * and apicid_to_node[] tables have valid entries for a CPU.
649 * This means we skip cpu_to_node[] initialisation for NUMA
650 * emulation and faking node case (when running a kernel compiled
651 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
652 * is already initialized in a round robin manner at numa_init_array,
653 * prior to this call, and this initialization is good enough
654 * for the fake NUMA cases.
656 void __init init_cpu_to_node(void)
660 for (i = 0; i < NR_CPUS; i++) {
662 u16 apicid = x86_cpu_to_apicid_init[i];
664 if (apicid == BAD_APICID)
666 node = apicid_to_node[apicid];
667 if (node == NUMA_NO_NODE)
669 if (!node_online(node))
671 numa_set_node(i, node);