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)
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)
79 memnodemap[addr >> shift] = i;
80 addr += (1UL << shift);
87 static int __init allocate_cachealigned_memnodemap(void)
91 memnodemap = memnode.embedded_map;
92 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
96 nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
97 nodemap_addr = find_e820_area(addr, end_pfn<<PAGE_SHIFT,
98 nodemap_size, L1_CACHE_BYTES);
99 if (nodemap_addr == -1UL) {
101 "NUMA: Unable to allocate Memory to Node hash map\n");
102 nodemap_addr = nodemap_size = 0;
105 memnodemap = phys_to_virt(nodemap_addr);
106 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
108 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
109 nodemap_addr, nodemap_addr + nodemap_size);
114 * The LSB of all start and end addresses in the node map is the value of the
115 * maximum possible shift.
117 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
120 int i, nodes_used = 0;
121 unsigned long start, end;
122 unsigned long bitfield = 0, memtop = 0;
124 for (i = 0; i < numnodes; i++) {
125 start = nodes[i].start;
137 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
138 memnodemapsize = (memtop >> i)+1;
142 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
146 shift = extract_lsb_from_nodes(nodes, numnodes);
147 if (allocate_cachealigned_memnodemap())
149 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
152 if (populate_memnodemap(nodes, numnodes, shift) != 1) {
153 printk(KERN_INFO "Your memory is not aligned you need to "
154 "rebuild your kernel with a bigger NODEMAPSIZE "
155 "shift=%d\n", shift);
161 int early_pfn_to_nid(unsigned long pfn)
163 return phys_to_nid(pfn << PAGE_SHIFT);
166 static void * __init early_node_mem(int nodeid, unsigned long start,
167 unsigned long end, unsigned long size,
170 unsigned long mem = find_e820_area(start, end, size, align);
176 ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
178 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
185 /* Initialize bootmem allocator for a node */
186 void __init setup_node_bootmem(int nodeid, unsigned long start,
189 unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
190 unsigned long bootmap_start, nodedata_phys;
192 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
194 start = round_up(start, ZONE_ALIGN);
196 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
199 start_pfn = start >> PAGE_SHIFT;
200 end_pfn = end >> PAGE_SHIFT;
202 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
204 if (node_data[nodeid] == NULL)
206 nodedata_phys = __pa(node_data[nodeid]);
207 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
208 nodedata_phys + pgdat_size - 1);
210 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
211 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
212 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
213 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
215 /* Find a place for the bootmem map */
216 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
217 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
219 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
220 * to use that to align to PAGE_SIZE
222 bootmap = early_node_mem(nodeid, bootmap_start, end,
223 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
224 if (bootmap == NULL) {
225 if (nodedata_phys < start || nodedata_phys >= end)
226 free_bootmem(nodedata_phys, pgdat_size);
227 node_data[nodeid] = NULL;
230 bootmap_start = __pa(bootmap);
232 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
233 bootmap_start >> PAGE_SHIFT,
236 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
237 bootmap_start, bootmap_start + bootmap_size - 1,
240 free_bootmem_with_active_regions(nodeid, end);
242 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size,
244 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
245 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
246 #ifdef CONFIG_ACPI_NUMA
247 srat_reserve_add_area(nodeid);
249 node_set_online(nodeid);
253 * There are unfortunately some poorly designed mainboards around that
254 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
255 * mapping. To avoid this fill in the mapping for all possible CPUs,
256 * as the number of CPUs is not known yet. We round robin the existing
259 void __init numa_init_array(void)
263 rr = first_node(node_online_map);
264 for (i = 0; i < NR_CPUS; i++) {
265 if (early_cpu_to_node(i) != NUMA_NO_NODE)
267 numa_set_node(i, rr);
268 rr = next_node(rr, node_online_map);
269 if (rr == MAX_NUMNODES)
270 rr = first_node(node_online_map);
274 #ifdef CONFIG_NUMA_EMU
276 char *cmdline __initdata;
279 * Setups up nid to range from addr to addr + size. If the end
280 * boundary is greater than max_addr, then max_addr is used instead.
281 * The return value is 0 if there is additional memory left for
282 * allocation past addr and -1 otherwise. addr is adjusted to be at
283 * the end of the node.
285 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
286 u64 size, u64 max_addr)
290 nodes[nid].start = *addr;
292 if (*addr >= max_addr) {
296 nodes[nid].end = *addr;
297 node_set(nid, node_possible_map);
298 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
299 nodes[nid].start, nodes[nid].end,
300 (nodes[nid].end - nodes[nid].start) >> 20);
305 * Splits num_nodes nodes up equally starting at node_start. The return value
306 * is the number of nodes split up and addr is adjusted to be at the end of the
307 * last node allocated.
309 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
310 u64 max_addr, int node_start,
319 if (num_nodes > MAX_NUMNODES)
320 num_nodes = MAX_NUMNODES;
321 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
324 * Calculate the number of big nodes that can be allocated as a result
325 * of consolidating the leftovers.
327 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
330 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
331 size &= FAKE_NODE_MIN_HASH_MASK;
333 printk(KERN_ERR "Not enough memory for each node. "
334 "NUMA emulation disabled.\n");
338 for (i = node_start; i < num_nodes + node_start; i++) {
339 u64 end = *addr + size;
342 end += FAKE_NODE_MIN_SIZE;
344 * The final node can have the remaining system RAM. Other
345 * nodes receive roughly the same amount of available pages.
347 if (i == num_nodes + node_start - 1)
350 while (end - *addr - e820_hole_size(*addr, end) <
352 end += FAKE_NODE_MIN_SIZE;
353 if (end > max_addr) {
358 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
361 return i - node_start + 1;
365 * Splits the remaining system RAM into chunks of size. The remaining memory is
366 * always assigned to a final node and can be asymmetric. Returns the number of
369 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
370 u64 max_addr, int node_start, u64 size)
373 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
374 while (!setup_node_range(i++, nodes, addr, size, max_addr))
376 return i - node_start;
380 * Sets up the system RAM area from start_pfn to end_pfn according to the
381 * numa=fake command-line option.
383 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
385 struct bootnode nodes[MAX_NUMNODES];
386 u64 size, addr = start_pfn << PAGE_SHIFT;
387 u64 max_addr = end_pfn << PAGE_SHIFT;
388 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
390 memset(&nodes, 0, sizeof(nodes));
392 * If the numa=fake command-line is just a single number N, split the
393 * system RAM into N fake nodes.
395 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
396 long n = simple_strtol(cmdline, NULL, 0);
398 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
404 /* Parse the command line. */
405 for (coeff_flag = 0; ; cmdline++) {
406 if (*cmdline && isdigit(*cmdline)) {
407 num = num * 10 + *cmdline - '0';
410 if (*cmdline == '*') {
415 if (!*cmdline || *cmdline == ',') {
419 * Round down to the nearest FAKE_NODE_MIN_SIZE.
420 * Command-line coefficients are in megabytes.
422 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
424 for (i = 0; i < coeff; i++, num_nodes++)
425 if (setup_node_range(num_nodes, nodes,
426 &addr, size, max_addr) < 0)
438 /* Fill remainder of system RAM, if appropriate. */
439 if (addr < max_addr) {
440 if (coeff_flag && coeff < 0) {
441 /* Split remaining nodes into num-sized chunks */
442 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
446 switch (*(cmdline - 1)) {
448 /* Split remaining nodes into coeff chunks */
451 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
455 /* Do not allocate remaining system RAM */
458 /* Give one final node */
459 setup_node_range(num_nodes, nodes, &addr,
460 max_addr - addr, max_addr);
465 memnode_shift = compute_hash_shift(nodes, num_nodes);
466 if (memnode_shift < 0) {
468 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
474 * We need to vacate all active ranges that may have been registered by
475 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
476 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
478 remove_all_active_ranges();
479 #ifdef CONFIG_ACPI_NUMA
482 for_each_node_mask(i, node_possible_map) {
483 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
484 nodes[i].end >> PAGE_SHIFT);
485 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
487 acpi_fake_nodes(nodes, num_nodes);
491 #endif /* CONFIG_NUMA_EMU */
493 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
497 nodes_clear(node_possible_map);
498 nodes_clear(node_online_map);
500 #ifdef CONFIG_NUMA_EMU
501 if (cmdline && !numa_emulation(start_pfn, end_pfn))
503 nodes_clear(node_possible_map);
504 nodes_clear(node_online_map);
507 #ifdef CONFIG_ACPI_NUMA
508 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
509 end_pfn << PAGE_SHIFT))
511 nodes_clear(node_possible_map);
512 nodes_clear(node_online_map);
515 #ifdef CONFIG_K8_NUMA
516 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
517 end_pfn<<PAGE_SHIFT))
519 nodes_clear(node_possible_map);
520 nodes_clear(node_online_map);
522 printk(KERN_INFO "%s\n",
523 numa_off ? "NUMA turned off" : "No NUMA configuration found");
525 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
526 start_pfn << PAGE_SHIFT,
527 end_pfn << PAGE_SHIFT);
528 /* setup dummy node covering all memory */
530 memnodemap = memnode.embedded_map;
533 node_set(0, node_possible_map);
534 for (i = 0; i < NR_CPUS; i++)
536 /* cpumask_of_cpu() may not be available during early startup */
537 memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
538 cpu_set(0, node_to_cpumask_map[0]);
539 e820_register_active_regions(0, start_pfn, end_pfn);
540 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
543 __cpuinit void numa_add_cpu(int cpu)
546 (unsigned long *)&node_to_cpumask_map[early_cpu_to_node(cpu)]);
549 void __cpuinit numa_set_node(int cpu, int node)
551 int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
554 cpu_to_node_map[cpu] = node;
555 else if(per_cpu_offset(cpu))
556 per_cpu(x86_cpu_to_node_map, cpu) = node;
558 Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
561 unsigned long __init numa_free_all_bootmem(void)
563 unsigned long pages = 0;
566 for_each_online_node(i)
567 pages += free_all_bootmem_node(NODE_DATA(i));
572 void __init paging_init(void)
574 unsigned long max_zone_pfns[MAX_NR_ZONES];
576 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
577 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
578 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
579 max_zone_pfns[ZONE_NORMAL] = end_pfn;
581 sparse_memory_present_with_active_regions(MAX_NUMNODES);
584 free_area_init_nodes(max_zone_pfns);
587 static __init int numa_setup(char *opt)
591 if (!strncmp(opt, "off", 3))
593 #ifdef CONFIG_NUMA_EMU
594 if (!strncmp(opt, "fake=", 5))
597 #ifdef CONFIG_ACPI_NUMA
598 if (!strncmp(opt, "noacpi", 6))
600 if (!strncmp(opt, "hotadd=", 7))
601 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
605 early_param("numa", numa_setup);
608 * Setup early cpu_to_node.
610 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
611 * and apicid_to_node[] tables have valid entries for a CPU.
612 * This means we skip cpu_to_node[] initialisation for NUMA
613 * emulation and faking node case (when running a kernel compiled
614 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
615 * is already initialized in a round robin manner at numa_init_array,
616 * prior to this call, and this initialization is good enough
617 * for the fake NUMA cases.
619 void __init init_cpu_to_node(void)
623 for (i = 0; i < NR_CPUS; i++) {
625 u16 apicid = x86_cpu_to_apicid_init[i];
627 if (apicid == BAD_APICID)
629 node = apicid_to_node[apicid];
630 if (node == NUMA_NO_NODE)
632 if (!node_online(node))
634 numa_set_node(i, node);