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;
34 int x86_cpu_to_node_map_init[NR_CPUS] = {
35 [0 ... NR_CPUS-1] = NUMA_NO_NODE
37 void *x86_cpu_to_node_map_early_ptr;
38 DEFINE_PER_CPU(int, x86_cpu_to_node_map) = NUMA_NO_NODE;
39 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map);
40 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr);
42 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
43 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
46 cpumask_t node_to_cpumask_map[MAX_NUMNODES] __read_mostly;
47 EXPORT_SYMBOL(node_to_cpumask_map);
49 int numa_off __initdata;
50 unsigned long __initdata nodemap_addr;
51 unsigned long __initdata nodemap_size;
54 * Given a shift value, try to populate memnodemap[]
57 * 0 if memnodmap[] too small (of shift too small)
58 * -1 if node overlap or lost ram (shift too big)
60 static int __init populate_memnodemap(const struct bootnode *nodes,
61 int numnodes, int shift)
63 unsigned long addr, end;
66 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
67 for (i = 0; i < numnodes; i++) {
68 addr = nodes[i].start;
72 if ((end >> shift) >= memnodemapsize)
75 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
77 memnodemap[addr >> shift] = i;
78 addr += (1UL << shift);
85 static int __init allocate_cachealigned_memnodemap(void)
89 memnodemap = memnode.embedded_map;
90 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
94 nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
95 nodemap_addr = find_e820_area(addr, end_pfn<<PAGE_SHIFT,
96 nodemap_size, L1_CACHE_BYTES);
97 if (nodemap_addr == -1UL) {
99 "NUMA: Unable to allocate Memory to Node hash map\n");
100 nodemap_addr = nodemap_size = 0;
103 memnodemap = phys_to_virt(nodemap_addr);
104 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
106 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
107 nodemap_addr, nodemap_addr + nodemap_size);
112 * The LSB of all start and end addresses in the node map is the value of the
113 * maximum possible shift.
115 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
118 int i, nodes_used = 0;
119 unsigned long start, end;
120 unsigned long bitfield = 0, memtop = 0;
122 for (i = 0; i < numnodes; i++) {
123 start = nodes[i].start;
135 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
136 memnodemapsize = (memtop >> i)+1;
140 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
144 shift = extract_lsb_from_nodes(nodes, numnodes);
145 if (allocate_cachealigned_memnodemap())
147 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
150 if (populate_memnodemap(nodes, numnodes, shift) != 1) {
151 printk(KERN_INFO "Your memory is not aligned you need to "
152 "rebuild your kernel with a bigger NODEMAPSIZE "
153 "shift=%d\n", shift);
159 int early_pfn_to_nid(unsigned long pfn)
161 return phys_to_nid(pfn << PAGE_SHIFT);
164 static void * __init early_node_mem(int nodeid, unsigned long start,
165 unsigned long end, unsigned long size,
168 unsigned long mem = find_e820_area(start, end, size, align);
174 ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
176 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
183 /* Initialize bootmem allocator for a node */
184 void __init setup_node_bootmem(int nodeid, unsigned long start,
187 unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
188 unsigned long bootmap_start, nodedata_phys;
190 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
192 start = round_up(start, ZONE_ALIGN);
194 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
197 start_pfn = start >> PAGE_SHIFT;
198 end_pfn = end >> PAGE_SHIFT;
200 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
202 if (node_data[nodeid] == NULL)
204 nodedata_phys = __pa(node_data[nodeid]);
205 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
206 nodedata_phys + pgdat_size - 1);
208 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
209 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
210 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
211 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
213 /* Find a place for the bootmem map */
214 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
215 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
217 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
218 * to use that to align to PAGE_SIZE
220 bootmap = early_node_mem(nodeid, bootmap_start, end,
221 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
222 if (bootmap == NULL) {
223 if (nodedata_phys < start || nodedata_phys >= end)
224 free_bootmem((unsigned long)node_data[nodeid],
226 node_data[nodeid] = NULL;
229 bootmap_start = __pa(bootmap);
231 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
232 bootmap_start >> PAGE_SHIFT,
235 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
236 bootmap_start, bootmap_start + bootmap_size - 1,
239 free_bootmem_with_active_regions(nodeid, end);
241 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size,
243 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
244 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
245 #ifdef CONFIG_ACPI_NUMA
246 srat_reserve_add_area(nodeid);
248 node_set_online(nodeid);
252 * There are unfortunately some poorly designed mainboards around that
253 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
254 * mapping. To avoid this fill in the mapping for all possible CPUs,
255 * as the number of CPUs is not known yet. We round robin the existing
258 void __init numa_init_array(void)
262 rr = first_node(node_online_map);
263 for (i = 0; i < NR_CPUS; i++) {
264 if (early_cpu_to_node(i) != NUMA_NO_NODE)
266 numa_set_node(i, rr);
267 rr = next_node(rr, node_online_map);
268 if (rr == MAX_NUMNODES)
269 rr = first_node(node_online_map);
273 #ifdef CONFIG_NUMA_EMU
275 char *cmdline __initdata;
278 * Setups up nid to range from addr to addr + size. If the end
279 * boundary is greater than max_addr, then max_addr is used instead.
280 * The return value is 0 if there is additional memory left for
281 * allocation past addr and -1 otherwise. addr is adjusted to be at
282 * the end of the node.
284 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
285 u64 size, u64 max_addr)
289 nodes[nid].start = *addr;
291 if (*addr >= max_addr) {
295 nodes[nid].end = *addr;
296 node_set(nid, node_possible_map);
297 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
298 nodes[nid].start, nodes[nid].end,
299 (nodes[nid].end - nodes[nid].start) >> 20);
304 * Splits num_nodes nodes up equally starting at node_start. The return value
305 * is the number of nodes split up and addr is adjusted to be at the end of the
306 * last node allocated.
308 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
309 u64 max_addr, int node_start,
318 if (num_nodes > MAX_NUMNODES)
319 num_nodes = MAX_NUMNODES;
320 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
323 * Calculate the number of big nodes that can be allocated as a result
324 * of consolidating the leftovers.
326 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
329 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
330 size &= FAKE_NODE_MIN_HASH_MASK;
332 printk(KERN_ERR "Not enough memory for each node. "
333 "NUMA emulation disabled.\n");
337 for (i = node_start; i < num_nodes + node_start; i++) {
338 u64 end = *addr + size;
341 end += FAKE_NODE_MIN_SIZE;
343 * The final node can have the remaining system RAM. Other
344 * nodes receive roughly the same amount of available pages.
346 if (i == num_nodes + node_start - 1)
349 while (end - *addr - e820_hole_size(*addr, end) <
351 end += FAKE_NODE_MIN_SIZE;
352 if (end > max_addr) {
357 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
360 return i - node_start + 1;
364 * Splits the remaining system RAM into chunks of size. The remaining memory is
365 * always assigned to a final node and can be asymmetric. Returns the number of
368 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
369 u64 max_addr, int node_start, u64 size)
372 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
373 while (!setup_node_range(i++, nodes, addr, size, max_addr))
375 return i - node_start;
379 * Sets up the system RAM area from start_pfn to end_pfn according to the
380 * numa=fake command-line option.
382 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
384 struct bootnode nodes[MAX_NUMNODES];
385 u64 size, addr = start_pfn << PAGE_SHIFT;
386 u64 max_addr = end_pfn << PAGE_SHIFT;
387 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
389 memset(&nodes, 0, sizeof(nodes));
391 * If the numa=fake command-line is just a single number N, split the
392 * system RAM into N fake nodes.
394 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
395 long n = simple_strtol(cmdline, NULL, 0);
397 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
403 /* Parse the command line. */
404 for (coeff_flag = 0; ; cmdline++) {
405 if (*cmdline && isdigit(*cmdline)) {
406 num = num * 10 + *cmdline - '0';
409 if (*cmdline == '*') {
414 if (!*cmdline || *cmdline == ',') {
418 * Round down to the nearest FAKE_NODE_MIN_SIZE.
419 * Command-line coefficients are in megabytes.
421 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
423 for (i = 0; i < coeff; i++, num_nodes++)
424 if (setup_node_range(num_nodes, nodes,
425 &addr, size, max_addr) < 0)
437 /* Fill remainder of system RAM, if appropriate. */
438 if (addr < max_addr) {
439 if (coeff_flag && coeff < 0) {
440 /* Split remaining nodes into num-sized chunks */
441 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
445 switch (*(cmdline - 1)) {
447 /* Split remaining nodes into coeff chunks */
450 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
454 /* Do not allocate remaining system RAM */
457 /* Give one final node */
458 setup_node_range(num_nodes, nodes, &addr,
459 max_addr - addr, max_addr);
464 memnode_shift = compute_hash_shift(nodes, num_nodes);
465 if (memnode_shift < 0) {
467 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
473 * We need to vacate all active ranges that may have been registered by
474 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
475 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
477 remove_all_active_ranges();
478 #ifdef CONFIG_ACPI_NUMA
481 for_each_node_mask(i, node_possible_map) {
482 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
483 nodes[i].end >> PAGE_SHIFT);
484 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
486 acpi_fake_nodes(nodes, num_nodes);
490 #endif /* CONFIG_NUMA_EMU */
492 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
496 nodes_clear(node_possible_map);
497 nodes_clear(node_online_map);
499 #ifdef CONFIG_NUMA_EMU
500 if (cmdline && !numa_emulation(start_pfn, end_pfn))
502 nodes_clear(node_possible_map);
503 nodes_clear(node_online_map);
506 #ifdef CONFIG_ACPI_NUMA
507 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
508 end_pfn << PAGE_SHIFT))
510 nodes_clear(node_possible_map);
511 nodes_clear(node_online_map);
514 #ifdef CONFIG_K8_NUMA
515 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
516 end_pfn<<PAGE_SHIFT))
518 nodes_clear(node_possible_map);
519 nodes_clear(node_online_map);
521 printk(KERN_INFO "%s\n",
522 numa_off ? "NUMA turned off" : "No NUMA configuration found");
524 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
525 start_pfn << PAGE_SHIFT,
526 end_pfn << PAGE_SHIFT);
527 /* setup dummy node covering all memory */
529 memnodemap = memnode.embedded_map;
532 node_set(0, node_possible_map);
533 for (i = 0; i < NR_CPUS; i++)
535 /* cpumask_of_cpu() may not be available during early startup */
536 memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
537 cpu_set(0, node_to_cpumask_map[0]);
538 e820_register_active_regions(0, start_pfn, end_pfn);
539 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
542 __cpuinit void numa_add_cpu(int cpu)
545 (unsigned long *)&node_to_cpumask_map[early_cpu_to_node(cpu)]);
548 void __cpuinit numa_set_node(int cpu, int node)
550 int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
552 cpu_pda(cpu)->nodenumber = node;
555 cpu_to_node_map[cpu] = node;
556 else if(per_cpu_offset(cpu))
557 per_cpu(x86_cpu_to_node_map, cpu) = node;
559 Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
562 unsigned long __init numa_free_all_bootmem(void)
564 unsigned long pages = 0;
567 for_each_online_node(i)
568 pages += free_all_bootmem_node(NODE_DATA(i));
573 void __init paging_init(void)
575 unsigned long max_zone_pfns[MAX_NR_ZONES];
577 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
578 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
579 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
580 max_zone_pfns[ZONE_NORMAL] = end_pfn;
582 sparse_memory_present_with_active_regions(MAX_NUMNODES);
585 free_area_init_nodes(max_zone_pfns);
588 static __init int numa_setup(char *opt)
592 if (!strncmp(opt, "off", 3))
594 #ifdef CONFIG_NUMA_EMU
595 if (!strncmp(opt, "fake=", 5))
598 #ifdef CONFIG_ACPI_NUMA
599 if (!strncmp(opt, "noacpi", 6))
601 if (!strncmp(opt, "hotadd=", 7))
602 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
606 early_param("numa", numa_setup);
609 * Setup early cpu_to_node.
611 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
612 * and apicid_to_node[] tables have valid entries for a CPU.
613 * This means we skip cpu_to_node[] initialisation for NUMA
614 * emulation and faking node case (when running a kernel compiled
615 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
616 * is already initialized in a round robin manner at numa_init_array,
617 * prior to this call, and this initialization is good enough
618 * for the fake NUMA cases.
620 void __init init_cpu_to_node(void)
624 for (i = 0; i < NR_CPUS; i++) {
625 u16 apicid = x86_cpu_to_apicid_init[i];
627 if (apicid == BAD_APICID)
629 if (apicid_to_node[apicid] == NUMA_NO_NODE)
631 numa_set_node(i, apicid_to_node[apicid]);