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 static __init int setup_node(int pxm)
41 return acpi_map_pxm_to_node(pxm);
44 static __init int conflicting_memblks(unsigned long start, unsigned long end)
47 for (i = 0; i < num_node_memblks; i++) {
48 struct bootnode *nd = &node_memblk_range[i];
49 if (nd->start == nd->end)
51 if (nd->end > start && nd->start < end)
52 return memblk_nodeid[i];
53 if (nd->end == end && nd->start == start)
54 return memblk_nodeid[i];
59 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
61 struct bootnode *nd = &nodes[i];
63 if (nd->start < start) {
65 if (nd->end < nd->start)
70 if (nd->start > nd->end)
75 static __init void bad_srat(void)
78 printk(KERN_ERR "SRAT: SRAT not used.\n");
80 for (i = 0; i < MAX_LOCAL_APIC; i++)
81 apicid_to_node[i] = NUMA_NO_NODE;
82 for (i = 0; i < MAX_NUMNODES; i++)
83 nodes_add[i].start = nodes[i].end = 0;
84 remove_all_active_ranges();
87 static __init inline int srat_disabled(void)
89 return numa_off || acpi_numa < 0;
92 /* Callback for SLIT parsing */
93 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
98 length = slit->header.length;
99 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
103 panic(" Can not save slit!\n");
105 acpi_slit = __va(phys);
106 memcpy(acpi_slit, slit, length);
107 reserve_early(phys, phys + length, "ACPI SLIT");
110 /* Callback for Proximity Domain -> x2APIC mapping */
112 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
119 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
123 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
125 pxm = pa->proximity_domain;
126 node = setup_node(pxm);
128 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
133 apic_id = pa->apic_id;
134 apicid_to_node[apic_id] = node;
135 node_set(node, cpu_nodes_parsed);
137 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
141 /* Callback for Proximity Domain -> LAPIC mapping */
143 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
150 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
154 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
156 pxm = pa->proximity_domain_lo;
157 node = setup_node(pxm);
159 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
164 if (get_uv_system_type() >= UV_X2APIC)
165 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
167 apic_id = pa->apic_id;
168 apicid_to_node[apic_id] = node;
169 node_set(node, cpu_nodes_parsed);
171 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
175 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
176 static inline int save_add_info(void) {return 1;}
178 static inline int save_add_info(void) {return 0;}
182 * This code supports one contiguous hot add area per node
185 update_nodes_add(int node, unsigned long start, unsigned long end)
187 unsigned long s_pfn = start >> PAGE_SHIFT;
188 unsigned long e_pfn = end >> PAGE_SHIFT;
190 struct bootnode *nd = &nodes_add[node];
192 /* I had some trouble with strange memory hotadd regions breaking
193 the boot. Be very strict here and reject anything unexpected.
194 If you want working memory hotadd write correct SRATs.
196 The node size check is a basic sanity check to guard against
198 if ((signed long)(end - start) < NODE_MIN_SIZE) {
199 printk(KERN_ERR "SRAT: Hotplug area too small\n");
203 /* This check might be a bit too strict, but I'm keeping it for now. */
204 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
206 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
213 if (nd->start == nd->end) {
218 if (nd->start == end) {
222 if (nd->end == start) {
227 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
231 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n",
235 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
237 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
239 struct bootnode *nd, oldnode;
240 unsigned long start, end;
246 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
250 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
253 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
255 start = ma->base_address;
256 end = start + ma->length;
257 pxm = ma->proximity_domain;
258 node = setup_node(pxm);
260 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
264 i = conflicting_memblks(start, end);
267 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
268 pxm, start, end, nodes[i].start, nodes[i].end);
271 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
272 pxm, start, end, node_to_pxm(i),
273 nodes[i].start, nodes[i].end);
279 if (!node_test_and_set(node, nodes_parsed)) {
283 if (start < nd->start)
289 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
291 e820_register_active_regions(node, start >> PAGE_SHIFT,
294 if (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) {
295 update_nodes_add(node, start, end);
296 /* restore nodes[node] */
298 if ((nd->start | nd->end) == 0)
299 node_clear(node, nodes_parsed);
302 node_memblk_range[num_node_memblks].start = start;
303 node_memblk_range[num_node_memblks].end = end;
304 memblk_nodeid[num_node_memblks] = node;
308 /* Sanity check to catch more bad SRATs (they are amazingly common).
309 Make sure the PXMs cover all memory. */
310 static int __init nodes_cover_memory(const struct bootnode *nodes)
313 unsigned long pxmram, e820ram;
316 for_each_node_mask(i, nodes_parsed) {
317 unsigned long s = nodes[i].start >> PAGE_SHIFT;
318 unsigned long e = nodes[i].end >> PAGE_SHIFT;
320 pxmram -= absent_pages_in_range(s, e);
321 if ((long)pxmram < 0)
325 e820ram = max_pfn - (e820_hole_size(0, max_pfn<<PAGE_SHIFT)>>PAGE_SHIFT);
326 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
327 if ((long)(e820ram - pxmram) >= (1<<(20 - PAGE_SHIFT))) {
329 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
330 (pxmram << PAGE_SHIFT) >> 20,
331 (e820ram << PAGE_SHIFT) >> 20);
337 void __init acpi_numa_arch_fixup(void) {}
339 /* Use the information discovered above to actually set up the nodes. */
340 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
347 /* First clean up the node list */
348 for (i = 0; i < MAX_NUMNODES; i++)
349 cutoff_node(i, start, end);
351 if (!nodes_cover_memory(nodes)) {
356 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
358 if (memnode_shift < 0) {
360 "SRAT: No NUMA node hash function found. Contact maintainer\n");
365 /* Account for nodes with cpus and no memory */
366 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed);
368 /* Finally register nodes */
369 for_each_node_mask(i, node_possible_map)
370 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
371 /* Try again in case setup_node_bootmem missed one due
372 to missing bootmem */
373 for_each_node_mask(i, node_possible_map)
375 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
377 for (i = 0; i < nr_cpu_ids; i++) {
378 int node = early_cpu_to_node(i);
380 if (node == NUMA_NO_NODE)
382 if (!node_online(node))
389 #ifdef CONFIG_NUMA_EMU
390 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
391 [0 ... MAX_NUMNODES-1] = PXM_INVAL
393 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
394 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
396 static int __init find_node_by_addr(unsigned long addr)
398 int ret = NUMA_NO_NODE;
401 for_each_node_mask(i, nodes_parsed) {
403 * Find the real node that this emulated node appears on. For
404 * the sake of simplicity, we only use a real node's starting
405 * address to determine which emulated node it appears on.
407 if (addr >= nodes[i].start && addr < nodes[i].end) {
416 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
417 * mappings that respect the real ACPI topology but reflect our emulated
418 * environment. For each emulated node, we find which real node it appears on
419 * and create PXM to NID mappings for those fake nodes which mirror that
420 * locality. SLIT will now represent the correct distances between emulated
421 * nodes as a result of the real topology.
423 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
427 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
429 for (i = 0; i < num_nodes; i++) {
432 nid = find_node_by_addr(fake_nodes[i].start);
433 if (nid == NUMA_NO_NODE)
435 pxm = node_to_pxm(nid);
436 if (pxm == PXM_INVAL)
438 fake_node_to_pxm_map[i] = pxm;
440 * For each apicid_to_node mapping that exists for this real
441 * node, it must now point to the fake node ID.
443 for (j = 0; j < MAX_LOCAL_APIC; j++)
444 if (apicid_to_node[j] == nid)
445 fake_apicid_to_node[j] = i;
447 for (i = 0; i < num_nodes; i++)
448 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
449 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
451 nodes_clear(nodes_parsed);
452 for (i = 0; i < num_nodes; i++)
453 if (fake_nodes[i].start != fake_nodes[i].end)
454 node_set(i, nodes_parsed);
455 WARN_ON(!nodes_cover_memory(fake_nodes));
458 static int null_slit_node_compare(int a, int b)
460 return node_to_pxm(a) == node_to_pxm(b);
463 static int null_slit_node_compare(int a, int b)
467 #endif /* CONFIG_NUMA_EMU */
469 int __node_distance(int a, int b)
474 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
476 index = acpi_slit->locality_count * node_to_pxm(a);
477 return acpi_slit->entry[index + node_to_pxm(b)];
480 EXPORT_SYMBOL(__node_distance);
482 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
483 int memory_add_physaddr_to_nid(u64 start)
488 if (nodes_add[i].start <= start && nodes_add[i].end > start)
493 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);