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);
200 start = round_up(start, ZONE_ALIGN);
202 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
205 start_pfn = start >> PAGE_SHIFT;
206 end_pfn = end >> PAGE_SHIFT;
208 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
210 if (node_data[nodeid] == NULL)
212 nodedata_phys = __pa(node_data[nodeid]);
213 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
214 nodedata_phys + pgdat_size - 1);
216 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
217 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
218 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
219 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
221 /* Find a place for the bootmem map */
222 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
223 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
225 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
226 * to use that to align to PAGE_SIZE
228 bootmap = early_node_mem(nodeid, bootmap_start, end,
229 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
230 if (bootmap == NULL) {
231 if (nodedata_phys < start || nodedata_phys >= end)
232 free_bootmem(nodedata_phys, pgdat_size);
233 node_data[nodeid] = NULL;
236 bootmap_start = __pa(bootmap);
238 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
239 bootmap_start >> PAGE_SHIFT,
242 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
243 bootmap_start, bootmap_start + bootmap_size - 1,
246 free_bootmem_with_active_regions(nodeid, end);
248 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size,
250 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
251 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
252 #ifdef CONFIG_ACPI_NUMA
253 srat_reserve_add_area(nodeid);
255 node_set_online(nodeid);
259 * There are unfortunately some poorly designed mainboards around that
260 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
261 * mapping. To avoid this fill in the mapping for all possible CPUs,
262 * as the number of CPUs is not known yet. We round robin the existing
265 void __init numa_init_array(void)
269 rr = first_node(node_online_map);
270 for (i = 0; i < NR_CPUS; i++) {
271 if (early_cpu_to_node(i) != NUMA_NO_NODE)
273 numa_set_node(i, rr);
274 rr = next_node(rr, node_online_map);
275 if (rr == MAX_NUMNODES)
276 rr = first_node(node_online_map);
280 #ifdef CONFIG_NUMA_EMU
282 char *cmdline __initdata;
285 * Setups up nid to range from addr to addr + size. If the end
286 * boundary is greater than max_addr, then max_addr is used instead.
287 * The return value is 0 if there is additional memory left for
288 * allocation past addr and -1 otherwise. addr is adjusted to be at
289 * the end of the node.
291 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
292 u64 size, u64 max_addr)
296 nodes[nid].start = *addr;
298 if (*addr >= max_addr) {
302 nodes[nid].end = *addr;
303 node_set(nid, node_possible_map);
304 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
305 nodes[nid].start, nodes[nid].end,
306 (nodes[nid].end - nodes[nid].start) >> 20);
311 * Splits num_nodes nodes up equally starting at node_start. The return value
312 * is the number of nodes split up and addr is adjusted to be at the end of the
313 * last node allocated.
315 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
316 u64 max_addr, int node_start,
325 if (num_nodes > MAX_NUMNODES)
326 num_nodes = MAX_NUMNODES;
327 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
330 * Calculate the number of big nodes that can be allocated as a result
331 * of consolidating the leftovers.
333 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
336 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
337 size &= FAKE_NODE_MIN_HASH_MASK;
339 printk(KERN_ERR "Not enough memory for each node. "
340 "NUMA emulation disabled.\n");
344 for (i = node_start; i < num_nodes + node_start; i++) {
345 u64 end = *addr + size;
348 end += FAKE_NODE_MIN_SIZE;
350 * The final node can have the remaining system RAM. Other
351 * nodes receive roughly the same amount of available pages.
353 if (i == num_nodes + node_start - 1)
356 while (end - *addr - e820_hole_size(*addr, end) <
358 end += FAKE_NODE_MIN_SIZE;
359 if (end > max_addr) {
364 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
367 return i - node_start + 1;
371 * Splits the remaining system RAM into chunks of size. The remaining memory is
372 * always assigned to a final node and can be asymmetric. Returns the number of
375 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
376 u64 max_addr, int node_start, u64 size)
379 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
380 while (!setup_node_range(i++, nodes, addr, size, max_addr))
382 return i - node_start;
386 * Sets up the system RAM area from start_pfn to end_pfn according to the
387 * numa=fake command-line option.
389 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
391 struct bootnode nodes[MAX_NUMNODES];
392 u64 size, addr = start_pfn << PAGE_SHIFT;
393 u64 max_addr = end_pfn << PAGE_SHIFT;
394 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
396 memset(&nodes, 0, sizeof(nodes));
398 * If the numa=fake command-line is just a single number N, split the
399 * system RAM into N fake nodes.
401 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
402 long n = simple_strtol(cmdline, NULL, 0);
404 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
410 /* Parse the command line. */
411 for (coeff_flag = 0; ; cmdline++) {
412 if (*cmdline && isdigit(*cmdline)) {
413 num = num * 10 + *cmdline - '0';
416 if (*cmdline == '*') {
421 if (!*cmdline || *cmdline == ',') {
425 * Round down to the nearest FAKE_NODE_MIN_SIZE.
426 * Command-line coefficients are in megabytes.
428 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
430 for (i = 0; i < coeff; i++, num_nodes++)
431 if (setup_node_range(num_nodes, nodes,
432 &addr, size, max_addr) < 0)
444 /* Fill remainder of system RAM, if appropriate. */
445 if (addr < max_addr) {
446 if (coeff_flag && coeff < 0) {
447 /* Split remaining nodes into num-sized chunks */
448 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
452 switch (*(cmdline - 1)) {
454 /* Split remaining nodes into coeff chunks */
457 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
461 /* Do not allocate remaining system RAM */
464 /* Give one final node */
465 setup_node_range(num_nodes, nodes, &addr,
466 max_addr - addr, max_addr);
471 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
472 if (memnode_shift < 0) {
474 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
480 * We need to vacate all active ranges that may have been registered by
481 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
482 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
484 remove_all_active_ranges();
485 #ifdef CONFIG_ACPI_NUMA
488 for_each_node_mask(i, node_possible_map) {
489 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
490 nodes[i].end >> PAGE_SHIFT);
491 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
493 acpi_fake_nodes(nodes, num_nodes);
497 #endif /* CONFIG_NUMA_EMU */
499 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
503 nodes_clear(node_possible_map);
504 nodes_clear(node_online_map);
506 #ifdef CONFIG_NUMA_EMU
507 if (cmdline && !numa_emulation(start_pfn, end_pfn))
509 nodes_clear(node_possible_map);
510 nodes_clear(node_online_map);
513 #ifdef CONFIG_ACPI_NUMA
514 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
515 end_pfn << PAGE_SHIFT))
517 nodes_clear(node_possible_map);
518 nodes_clear(node_online_map);
521 #ifdef CONFIG_K8_NUMA
522 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
523 end_pfn<<PAGE_SHIFT))
525 nodes_clear(node_possible_map);
526 nodes_clear(node_online_map);
528 printk(KERN_INFO "%s\n",
529 numa_off ? "NUMA turned off" : "No NUMA configuration found");
531 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
532 start_pfn << PAGE_SHIFT,
533 end_pfn << PAGE_SHIFT);
534 /* setup dummy node covering all memory */
536 memnodemap = memnode.embedded_map;
539 node_set(0, node_possible_map);
540 for (i = 0; i < NR_CPUS; i++)
542 /* cpumask_of_cpu() may not be available during early startup */
543 memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
544 cpu_set(0, node_to_cpumask_map[0]);
545 e820_register_active_regions(0, start_pfn, end_pfn);
546 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
549 __cpuinit void numa_add_cpu(int cpu)
552 (unsigned long *)&node_to_cpumask_map[early_cpu_to_node(cpu)]);
555 void __cpuinit numa_set_node(int cpu, int node)
557 int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
560 cpu_to_node_map[cpu] = node;
561 else if(per_cpu_offset(cpu))
562 per_cpu(x86_cpu_to_node_map, cpu) = node;
564 Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
567 unsigned long __init numa_free_all_bootmem(void)
569 unsigned long pages = 0;
572 for_each_online_node(i)
573 pages += free_all_bootmem_node(NODE_DATA(i));
578 void __init paging_init(void)
580 unsigned long max_zone_pfns[MAX_NR_ZONES];
582 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
583 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
584 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
585 max_zone_pfns[ZONE_NORMAL] = end_pfn;
587 sparse_memory_present_with_active_regions(MAX_NUMNODES);
590 free_area_init_nodes(max_zone_pfns);
593 static __init int numa_setup(char *opt)
597 if (!strncmp(opt, "off", 3))
599 #ifdef CONFIG_NUMA_EMU
600 if (!strncmp(opt, "fake=", 5))
603 #ifdef CONFIG_ACPI_NUMA
604 if (!strncmp(opt, "noacpi", 6))
606 if (!strncmp(opt, "hotadd=", 7))
607 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
611 early_param("numa", numa_setup);
614 * Setup early cpu_to_node.
616 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
617 * and apicid_to_node[] tables have valid entries for a CPU.
618 * This means we skip cpu_to_node[] initialisation for NUMA
619 * emulation and faking node case (when running a kernel compiled
620 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
621 * is already initialized in a round robin manner at numa_init_array,
622 * prior to this call, and this initialization is good enough
623 * for the fake NUMA cases.
625 void __init init_cpu_to_node(void)
629 for (i = 0; i < NR_CPUS; i++) {
631 u16 apicid = x86_cpu_to_apicid_init[i];
633 if (apicid == BAD_APICID)
635 node = apicid_to_node[apicid];
636 if (node == NUMA_NO_NODE)
638 if (!node_online(node))
640 numa_set_node(i, node);