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 static bootmem_data_t plat_node_bdata[MAX_NUMNODES];
32 struct memnode memnode;
34 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
35 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
38 int numa_off __initdata;
39 static unsigned long __initdata nodemap_addr;
40 static unsigned long __initdata nodemap_size;
43 * Given a shift value, try to populate memnodemap[]
46 * 0 if memnodmap[] too small (of shift too small)
47 * -1 if node overlap or lost ram (shift too big)
49 static int __init populate_memnodemap(const struct bootnode *nodes,
50 int numnodes, int shift, int *nodeids)
52 unsigned long addr, end;
55 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
56 for (i = 0; i < numnodes; i++) {
57 addr = nodes[i].start;
61 if ((end >> shift) >= memnodemapsize)
64 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
68 memnodemap[addr >> shift] = i;
70 memnodemap[addr >> shift] = nodeids[i];
72 addr += (1UL << shift);
79 static int __init allocate_cachealigned_memnodemap(void)
83 memnodemap = memnode.embedded_map;
84 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
88 nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
89 nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT,
90 nodemap_size, L1_CACHE_BYTES);
91 if (nodemap_addr == -1UL) {
93 "NUMA: Unable to allocate Memory to Node hash map\n");
94 nodemap_addr = nodemap_size = 0;
97 memnodemap = phys_to_virt(nodemap_addr);
98 reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
100 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
101 nodemap_addr, nodemap_addr + nodemap_size);
106 * The LSB of all start and end addresses in the node map is the value of the
107 * maximum possible shift.
109 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
112 int i, nodes_used = 0;
113 unsigned long start, end;
114 unsigned long bitfield = 0, memtop = 0;
116 for (i = 0; i < numnodes; i++) {
117 start = nodes[i].start;
129 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
130 memnodemapsize = (memtop >> i)+1;
134 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
139 shift = extract_lsb_from_nodes(nodes, numnodes);
140 if (allocate_cachealigned_memnodemap())
142 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
145 if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
146 printk(KERN_INFO "Your memory is not aligned you need to "
147 "rebuild your kernel with a bigger NODEMAPSIZE "
148 "shift=%d\n", shift);
154 int early_pfn_to_nid(unsigned long pfn)
156 return phys_to_nid(pfn << PAGE_SHIFT);
159 static void * __init early_node_mem(int nodeid, unsigned long start,
160 unsigned long end, unsigned long size,
163 unsigned long mem = find_e820_area(start, end, size, align);
169 ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
171 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
178 /* Initialize bootmem allocator for a node */
179 void __init setup_node_bootmem(int nodeid, unsigned long start,
182 unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size;
183 unsigned long bootmap_start, nodedata_phys;
185 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
188 start = round_up(start, ZONE_ALIGN);
190 printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
193 start_pfn = start >> PAGE_SHIFT;
194 last_pfn = end >> PAGE_SHIFT;
196 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
198 if (node_data[nodeid] == NULL)
200 nodedata_phys = __pa(node_data[nodeid]);
201 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
202 nodedata_phys + pgdat_size - 1);
204 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
205 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
206 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
207 NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
210 * Find a place for the bootmem map
211 * nodedata_phys could be on other nodes by alloc_bootmem,
212 * so need to sure bootmap_start not to be small, otherwise
213 * early_node_mem will get that with find_e820_area instead
214 * of alloc_bootmem, that could clash with reserved range
216 bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn);
217 nid = phys_to_nid(nodedata_phys);
219 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
221 bootmap_start = round_up(start, PAGE_SIZE);
223 * SMP_CACHE_BYTES could be enough, but init_bootmem_node like
224 * to use that to align to PAGE_SIZE
226 bootmap = early_node_mem(nodeid, bootmap_start, end,
227 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
228 if (bootmap == NULL) {
229 if (nodedata_phys < start || nodedata_phys >= end)
230 free_bootmem(nodedata_phys, pgdat_size);
231 node_data[nodeid] = NULL;
234 bootmap_start = __pa(bootmap);
236 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
237 bootmap_start >> PAGE_SHIFT,
238 start_pfn, last_pfn);
240 printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
241 bootmap_start, bootmap_start + bootmap_size - 1,
244 free_bootmem_with_active_regions(nodeid, end);
247 * convert early reserve to bootmem reserve earlier
248 * otherwise early_node_mem could use early reserved mem
251 early_res_to_bootmem(start, end);
254 * in some case early_node_mem could use alloc_bootmem
255 * to get range on other node, don't reserve that again
258 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
260 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys,
261 pgdat_size, BOOTMEM_DEFAULT);
262 nid = phys_to_nid(bootmap_start);
264 printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid);
266 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
267 bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
269 #ifdef CONFIG_ACPI_NUMA
270 srat_reserve_add_area(nodeid);
272 node_set_online(nodeid);
276 * There are unfortunately some poorly designed mainboards around that
277 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
278 * mapping. To avoid this fill in the mapping for all possible CPUs,
279 * as the number of CPUs is not known yet. We round robin the existing
282 void __init numa_init_array(void)
286 rr = first_node(node_online_map);
287 for (i = 0; i < NR_CPUS; i++) {
288 if (early_cpu_to_node(i) != NUMA_NO_NODE)
290 numa_set_node(i, rr);
291 rr = next_node(rr, node_online_map);
292 if (rr == MAX_NUMNODES)
293 rr = first_node(node_online_map);
297 #ifdef CONFIG_NUMA_EMU
299 static char *cmdline __initdata;
302 * Setups up nid to range from addr to addr + size. If the end
303 * boundary is greater than max_addr, then max_addr is used instead.
304 * The return value is 0 if there is additional memory left for
305 * allocation past addr and -1 otherwise. addr is adjusted to be at
306 * the end of the node.
308 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
309 u64 size, u64 max_addr)
313 nodes[nid].start = *addr;
315 if (*addr >= max_addr) {
319 nodes[nid].end = *addr;
320 node_set(nid, node_possible_map);
321 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
322 nodes[nid].start, nodes[nid].end,
323 (nodes[nid].end - nodes[nid].start) >> 20);
328 * Splits num_nodes nodes up equally starting at node_start. The return value
329 * is the number of nodes split up and addr is adjusted to be at the end of the
330 * last node allocated.
332 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
333 u64 max_addr, int node_start,
342 if (num_nodes > MAX_NUMNODES)
343 num_nodes = MAX_NUMNODES;
344 size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
347 * Calculate the number of big nodes that can be allocated as a result
348 * of consolidating the leftovers.
350 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
353 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
354 size &= FAKE_NODE_MIN_HASH_MASK;
356 printk(KERN_ERR "Not enough memory for each node. "
357 "NUMA emulation disabled.\n");
361 for (i = node_start; i < num_nodes + node_start; i++) {
362 u64 end = *addr + size;
365 end += FAKE_NODE_MIN_SIZE;
367 * The final node can have the remaining system RAM. Other
368 * nodes receive roughly the same amount of available pages.
370 if (i == num_nodes + node_start - 1)
373 while (end - *addr - e820_hole_size(*addr, end) <
375 end += FAKE_NODE_MIN_SIZE;
376 if (end > max_addr) {
381 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
384 return i - node_start + 1;
388 * Splits the remaining system RAM into chunks of size. The remaining memory is
389 * always assigned to a final node and can be asymmetric. Returns the number of
392 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
393 u64 max_addr, int node_start, u64 size)
396 size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
397 while (!setup_node_range(i++, nodes, addr, size, max_addr))
399 return i - node_start;
403 * Sets up the system RAM area from start_pfn to last_pfn according to the
404 * numa=fake command-line option.
406 static struct bootnode nodes[MAX_NUMNODES] __initdata;
408 static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn)
410 u64 size, addr = start_pfn << PAGE_SHIFT;
411 u64 max_addr = last_pfn << PAGE_SHIFT;
412 int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
414 memset(&nodes, 0, sizeof(nodes));
416 * If the numa=fake command-line is just a single number N, split the
417 * system RAM into N fake nodes.
419 if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
420 long n = simple_strtol(cmdline, NULL, 0);
422 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
428 /* Parse the command line. */
429 for (coeff_flag = 0; ; cmdline++) {
430 if (*cmdline && isdigit(*cmdline)) {
431 num = num * 10 + *cmdline - '0';
434 if (*cmdline == '*') {
439 if (!*cmdline || *cmdline == ',') {
443 * Round down to the nearest FAKE_NODE_MIN_SIZE.
444 * Command-line coefficients are in megabytes.
446 size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
448 for (i = 0; i < coeff; i++, num_nodes++)
449 if (setup_node_range(num_nodes, nodes,
450 &addr, size, max_addr) < 0)
462 /* Fill remainder of system RAM, if appropriate. */
463 if (addr < max_addr) {
464 if (coeff_flag && coeff < 0) {
465 /* Split remaining nodes into num-sized chunks */
466 num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
470 switch (*(cmdline - 1)) {
472 /* Split remaining nodes into coeff chunks */
475 num_nodes += split_nodes_equally(nodes, &addr, max_addr,
479 /* Do not allocate remaining system RAM */
482 /* Give one final node */
483 setup_node_range(num_nodes, nodes, &addr,
484 max_addr - addr, max_addr);
489 memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
490 if (memnode_shift < 0) {
492 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
498 * We need to vacate all active ranges that may have been registered by
499 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
500 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
502 remove_all_active_ranges();
503 #ifdef CONFIG_ACPI_NUMA
506 for_each_node_mask(i, node_possible_map) {
507 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
508 nodes[i].end >> PAGE_SHIFT);
509 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
511 acpi_fake_nodes(nodes, num_nodes);
515 #endif /* CONFIG_NUMA_EMU */
517 void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn)
521 nodes_clear(node_possible_map);
522 nodes_clear(node_online_map);
524 #ifdef CONFIG_NUMA_EMU
525 if (cmdline && !numa_emulation(start_pfn, last_pfn))
527 nodes_clear(node_possible_map);
528 nodes_clear(node_online_map);
531 #ifdef CONFIG_ACPI_NUMA
532 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
533 last_pfn << PAGE_SHIFT))
535 nodes_clear(node_possible_map);
536 nodes_clear(node_online_map);
539 #ifdef CONFIG_K8_NUMA
540 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
541 last_pfn<<PAGE_SHIFT))
543 nodes_clear(node_possible_map);
544 nodes_clear(node_online_map);
546 printk(KERN_INFO "%s\n",
547 numa_off ? "NUMA turned off" : "No NUMA configuration found");
549 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
550 start_pfn << PAGE_SHIFT,
551 last_pfn << PAGE_SHIFT);
552 /* setup dummy node covering all memory */
554 memnodemap = memnode.embedded_map;
557 node_set(0, node_possible_map);
558 for (i = 0; i < NR_CPUS; i++)
560 e820_register_active_regions(0, start_pfn, last_pfn);
561 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
564 unsigned long __init numa_free_all_bootmem(void)
566 unsigned long pages = 0;
569 for_each_online_node(i)
570 pages += free_all_bootmem_node(NODE_DATA(i));
575 void __init paging_init(void)
577 unsigned long max_zone_pfns[MAX_NR_ZONES];
579 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
580 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
581 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
582 max_zone_pfns[ZONE_NORMAL] = max_pfn;
584 sparse_memory_present_with_active_regions(MAX_NUMNODES);
587 free_area_init_nodes(max_zone_pfns);
590 static __init int numa_setup(char *opt)
594 if (!strncmp(opt, "off", 3))
596 #ifdef CONFIG_NUMA_EMU
597 if (!strncmp(opt, "fake=", 5))
600 #ifdef CONFIG_ACPI_NUMA
601 if (!strncmp(opt, "noacpi", 6))
603 if (!strncmp(opt, "hotadd=", 7))
604 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
608 early_param("numa", numa_setup);
612 * Setup early cpu_to_node.
614 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
615 * and apicid_to_node[] tables have valid entries for a CPU.
616 * This means we skip cpu_to_node[] initialisation for NUMA
617 * emulation and faking node case (when running a kernel compiled
618 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
619 * is already initialized in a round robin manner at numa_init_array,
620 * prior to this call, and this initialization is good enough
621 * for the fake NUMA cases.
623 * Called before the per_cpu areas are setup.
625 void __init init_cpu_to_node(void)
628 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
630 BUG_ON(cpu_to_apicid == NULL);
632 for_each_possible_cpu(cpu) {
634 u16 apicid = cpu_to_apicid[cpu];
636 if (apicid == BAD_APICID)
638 node = apicid_to_node[apicid];
639 if (node == NUMA_NO_NODE)
641 if (!node_online(node))
643 numa_set_node(cpu, node);