2 * ACPI 3.0 based NUMA setup
3 * Copyright 2004 Andi Kleen, SuSE Labs.
5 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
7 * Called from acpi_numa_init while reading the SRAT and SLIT tables.
8 * Assumes all memory regions belonging to a single proximity domain
9 * are in one chunk. Holes between them will be included in the node.
12 #include <linux/kernel.h>
13 #include <linux/acpi.h>
14 #include <linux/mmzone.h>
15 #include <linux/bitmap.h>
16 #include <linux/module.h>
17 #include <linux/topology.h>
18 #include <linux/bootmem.h>
20 #include <asm/proto.h>
24 #include <asm/uv/uv.h>
26 int acpi_numa __initdata;
28 static struct acpi_table_slit *acpi_slit;
30 static nodemask_t nodes_parsed __initdata;
31 static nodemask_t cpu_nodes_parsed __initdata;
32 static struct bootnode nodes[MAX_NUMNODES] __initdata;
33 static struct bootnode nodes_add[MAX_NUMNODES];
35 static int num_node_memblks __initdata;
36 static struct bootnode node_memblk_range[NR_NODE_MEMBLKS] __initdata;
37 static int memblk_nodeid[NR_NODE_MEMBLKS] __initdata;
39 /* Too small nodes confuse the VM badly. Usually they result
41 #define NODE_MIN_SIZE (4*1024*1024)
43 static __init int setup_node(int pxm)
45 return acpi_map_pxm_to_node(pxm);
48 static __init int conflicting_memblks(unsigned long start, unsigned long end)
51 for (i = 0; i < num_node_memblks; i++) {
52 struct bootnode *nd = &node_memblk_range[i];
53 if (nd->start == nd->end)
55 if (nd->end > start && nd->start < end)
56 return memblk_nodeid[i];
57 if (nd->end == end && nd->start == start)
58 return memblk_nodeid[i];
63 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
65 struct bootnode *nd = &nodes[i];
67 if (nd->start < start) {
69 if (nd->end < nd->start)
74 if (nd->start > nd->end)
79 static __init void bad_srat(void)
82 printk(KERN_ERR "SRAT: SRAT not used.\n");
84 for (i = 0; i < MAX_LOCAL_APIC; i++)
85 apicid_to_node[i] = NUMA_NO_NODE;
86 for (i = 0; i < MAX_NUMNODES; i++)
87 nodes_add[i].start = nodes[i].end = 0;
88 remove_all_active_ranges();
91 static __init inline int srat_disabled(void)
93 return numa_off || acpi_numa < 0;
96 /* Callback for SLIT parsing */
97 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
102 length = slit->header.length;
103 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
107 panic(" Can not save slit!\n");
109 acpi_slit = __va(phys);
110 memcpy(acpi_slit, slit, length);
111 reserve_early(phys, phys + length, "ACPI SLIT");
114 /* Callback for Proximity Domain -> x2APIC mapping */
116 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
123 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
127 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
129 pxm = pa->proximity_domain;
130 node = setup_node(pxm);
132 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
137 apic_id = pa->apic_id;
138 apicid_to_node[apic_id] = node;
139 node_set(node, cpu_nodes_parsed);
141 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
145 /* Callback for Proximity Domain -> LAPIC mapping */
147 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
154 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
158 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
160 pxm = pa->proximity_domain_lo;
161 node = setup_node(pxm);
163 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
168 if (get_uv_system_type() >= UV_X2APIC)
169 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
171 apic_id = pa->apic_id;
172 apicid_to_node[apic_id] = node;
173 node_set(node, cpu_nodes_parsed);
175 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
179 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
180 static inline int save_add_info(void) {return 1;}
182 static inline int save_add_info(void) {return 0;}
186 * This code supports one contiguous hot add area per node
189 update_nodes_add(int node, unsigned long start, unsigned long end)
191 unsigned long s_pfn = start >> PAGE_SHIFT;
192 unsigned long e_pfn = end >> PAGE_SHIFT;
194 struct bootnode *nd = &nodes_add[node];
196 /* I had some trouble with strange memory hotadd regions breaking
197 the boot. Be very strict here and reject anything unexpected.
198 If you want working memory hotadd write correct SRATs.
200 The node size check is a basic sanity check to guard against
202 if ((signed long)(end - start) < NODE_MIN_SIZE) {
203 printk(KERN_ERR "SRAT: Hotplug area too small\n");
207 /* This check might be a bit too strict, but I'm keeping it for now. */
208 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
210 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
217 if (nd->start == nd->end) {
222 if (nd->start == end) {
226 if (nd->end == start) {
231 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
235 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n",
239 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
241 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
243 struct bootnode *nd, oldnode;
244 unsigned long start, end;
250 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
254 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
257 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
259 start = ma->base_address;
260 end = start + ma->length;
261 pxm = ma->proximity_domain;
262 node = setup_node(pxm);
264 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
268 i = conflicting_memblks(start, end);
271 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
272 pxm, start, end, nodes[i].start, nodes[i].end);
275 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
276 pxm, start, end, node_to_pxm(i),
277 nodes[i].start, nodes[i].end);
283 if (!node_test_and_set(node, nodes_parsed)) {
287 if (start < nd->start)
293 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
295 e820_register_active_regions(node, start >> PAGE_SHIFT,
298 if (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) {
299 update_nodes_add(node, start, end);
300 /* restore nodes[node] */
302 if ((nd->start | nd->end) == 0)
303 node_clear(node, nodes_parsed);
306 node_memblk_range[num_node_memblks].start = start;
307 node_memblk_range[num_node_memblks].end = end;
308 memblk_nodeid[num_node_memblks] = node;
312 /* Sanity check to catch more bad SRATs (they are amazingly common).
313 Make sure the PXMs cover all memory. */
314 static int __init nodes_cover_memory(const struct bootnode *nodes)
317 unsigned long pxmram, e820ram;
320 for_each_node_mask(i, nodes_parsed) {
321 unsigned long s = nodes[i].start >> PAGE_SHIFT;
322 unsigned long e = nodes[i].end >> PAGE_SHIFT;
324 pxmram -= absent_pages_in_range(s, e);
325 if ((long)pxmram < 0)
329 e820ram = max_pfn - (e820_hole_size(0, max_pfn<<PAGE_SHIFT)>>PAGE_SHIFT);
330 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
331 if ((long)(e820ram - pxmram) >= (1<<(20 - PAGE_SHIFT))) {
333 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
334 (pxmram << PAGE_SHIFT) >> 20,
335 (e820ram << PAGE_SHIFT) >> 20);
341 static void __init unparse_node(int node)
344 node_clear(node, nodes_parsed);
345 node_clear(node, cpu_nodes_parsed);
346 for (i = 0; i < MAX_LOCAL_APIC; i++) {
347 if (apicid_to_node[i] == node)
348 apicid_to_node[i] = NUMA_NO_NODE;
352 void __init acpi_numa_arch_fixup(void) {}
354 /* Use the information discovered above to actually set up the nodes. */
355 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
362 /* First clean up the node list */
363 for (i = 0; i < MAX_NUMNODES; i++) {
364 cutoff_node(i, start, end);
366 * don't confuse VM with a node that doesn't have the
370 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
376 if (!nodes_cover_memory(nodes)) {
381 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
383 if (memnode_shift < 0) {
385 "SRAT: No NUMA node hash function found. Contact maintainer\n");
390 /* Account for nodes with cpus and no memory */
391 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed);
393 /* Finally register nodes */
394 for_each_node_mask(i, node_possible_map)
395 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
396 /* Try again in case setup_node_bootmem missed one due
397 to missing bootmem */
398 for_each_node_mask(i, node_possible_map)
400 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
402 for (i = 0; i < nr_cpu_ids; i++) {
403 int node = early_cpu_to_node(i);
405 if (node == NUMA_NO_NODE)
407 if (!node_isset(node, node_possible_map))
414 #ifdef CONFIG_NUMA_EMU
415 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
416 [0 ... MAX_NUMNODES-1] = PXM_INVAL
418 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
419 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
421 static int __init find_node_by_addr(unsigned long addr)
423 int ret = NUMA_NO_NODE;
426 for_each_node_mask(i, nodes_parsed) {
428 * Find the real node that this emulated node appears on. For
429 * the sake of simplicity, we only use a real node's starting
430 * address to determine which emulated node it appears on.
432 if (addr >= nodes[i].start && addr < nodes[i].end) {
441 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
442 * mappings that respect the real ACPI topology but reflect our emulated
443 * environment. For each emulated node, we find which real node it appears on
444 * and create PXM to NID mappings for those fake nodes which mirror that
445 * locality. SLIT will now represent the correct distances between emulated
446 * nodes as a result of the real topology.
448 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
452 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
454 for (i = 0; i < num_nodes; i++) {
457 nid = find_node_by_addr(fake_nodes[i].start);
458 if (nid == NUMA_NO_NODE)
460 pxm = node_to_pxm(nid);
461 if (pxm == PXM_INVAL)
463 fake_node_to_pxm_map[i] = pxm;
465 * For each apicid_to_node mapping that exists for this real
466 * node, it must now point to the fake node ID.
468 for (j = 0; j < MAX_LOCAL_APIC; j++)
469 if (apicid_to_node[j] == nid)
470 fake_apicid_to_node[j] = i;
472 for (i = 0; i < num_nodes; i++)
473 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
474 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
476 nodes_clear(nodes_parsed);
477 for (i = 0; i < num_nodes; i++)
478 if (fake_nodes[i].start != fake_nodes[i].end)
479 node_set(i, nodes_parsed);
480 WARN_ON(!nodes_cover_memory(fake_nodes));
483 static int null_slit_node_compare(int a, int b)
485 return node_to_pxm(a) == node_to_pxm(b);
488 static int null_slit_node_compare(int a, int b)
492 #endif /* CONFIG_NUMA_EMU */
494 int __node_distance(int a, int b)
499 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
501 index = acpi_slit->locality_count * node_to_pxm(a);
502 return acpi_slit->entry[index + node_to_pxm(b)];
505 EXPORT_SYMBOL(__node_distance);
507 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
508 int memory_add_physaddr_to_nid(u64 start)
513 if (nodes_add[i].start <= start && nodes_add[i].end > start)
518 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);