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[i].start = nodes[i].end = 0;
84 nodes_add[i].start = nodes_add[i].end = 0;
86 remove_all_active_ranges();
89 static __init inline int srat_disabled(void)
91 return numa_off || acpi_numa < 0;
94 /* Callback for SLIT parsing */
95 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
100 length = slit->header.length;
101 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
105 panic(" Can not save slit!\n");
107 acpi_slit = __va(phys);
108 memcpy(acpi_slit, slit, length);
109 reserve_early(phys, phys + length, "ACPI SLIT");
112 /* Callback for Proximity Domain -> x2APIC mapping */
114 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
121 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
125 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
127 pxm = pa->proximity_domain;
128 node = setup_node(pxm);
130 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
135 apic_id = pa->apic_id;
136 apicid_to_node[apic_id] = node;
137 node_set(node, cpu_nodes_parsed);
139 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%04x -> Node %u\n",
143 /* Callback for Proximity Domain -> LAPIC mapping */
145 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
152 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
156 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
158 pxm = pa->proximity_domain_lo;
159 if (acpi_srat_revision >= 2)
160 pxm |= *((unsigned int*)pa->proximity_domain_hi) << 8;
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 0x%02x -> 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 node_set(node, cpu_nodes_parsed);
236 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n",
241 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
243 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
245 struct bootnode *nd, oldnode;
246 unsigned long start, end;
252 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
256 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
259 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
261 start = ma->base_address;
262 end = start + ma->length;
263 pxm = ma->proximity_domain;
264 if (acpi_srat_revision <= 1)
266 node = setup_node(pxm);
268 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
272 i = conflicting_memblks(start, end);
275 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
276 pxm, start, end, nodes[i].start, nodes[i].end);
279 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
280 pxm, start, end, node_to_pxm(i),
281 nodes[i].start, nodes[i].end);
287 if (!node_test_and_set(node, nodes_parsed)) {
291 if (start < nd->start)
297 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
300 if (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) {
301 update_nodes_add(node, start, end);
302 /* restore nodes[node] */
304 if ((nd->start | nd->end) == 0)
305 node_clear(node, nodes_parsed);
308 node_memblk_range[num_node_memblks].start = start;
309 node_memblk_range[num_node_memblks].end = end;
310 memblk_nodeid[num_node_memblks] = node;
314 /* Sanity check to catch more bad SRATs (they are amazingly common).
315 Make sure the PXMs cover all memory. */
316 static int __init nodes_cover_memory(const struct bootnode *nodes)
319 unsigned long pxmram, e820ram;
322 for_each_node_mask(i, nodes_parsed) {
323 unsigned long s = nodes[i].start >> PAGE_SHIFT;
324 unsigned long e = nodes[i].end >> PAGE_SHIFT;
326 pxmram -= __absent_pages_in_range(i, s, e);
327 if ((long)pxmram < 0)
331 e820ram = max_pfn - (e820_hole_size(0, max_pfn<<PAGE_SHIFT)>>PAGE_SHIFT);
332 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
333 if ((long)(e820ram - pxmram) >= (1<<(20 - PAGE_SHIFT))) {
335 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
336 (pxmram << PAGE_SHIFT) >> 20,
337 (e820ram << PAGE_SHIFT) >> 20);
343 void __init acpi_numa_arch_fixup(void) {}
345 int __init acpi_get_nodes(struct bootnode *physnodes)
350 for_each_node_mask(i, nodes_parsed) {
351 physnodes[ret].start = nodes[i].start;
352 physnodes[ret].end = nodes[i].end;
358 /* Use the information discovered above to actually set up the nodes. */
359 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
366 /* First clean up the node list */
367 for (i = 0; i < MAX_NUMNODES; i++)
368 cutoff_node(i, start, end);
370 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
372 if (memnode_shift < 0) {
374 "SRAT: No NUMA node hash function found. Contact maintainer\n");
379 for_each_node_mask(i, nodes_parsed)
380 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
381 nodes[i].end >> PAGE_SHIFT);
382 /* for out of order entries in SRAT */
384 if (!nodes_cover_memory(nodes)) {
389 /* Account for nodes with cpus and no memory */
390 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed);
392 /* Finally register nodes */
393 for_each_node_mask(i, node_possible_map)
394 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
395 /* Try again in case setup_node_bootmem missed one due
396 to missing bootmem */
397 for_each_node_mask(i, node_possible_map)
399 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
401 for (i = 0; i < nr_cpu_ids; i++) {
402 int node = early_cpu_to_node(i);
404 if (node == NUMA_NO_NODE)
406 if (!node_online(node))
413 #ifdef CONFIG_NUMA_EMU
414 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
415 [0 ... MAX_NUMNODES-1] = PXM_INVAL
417 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
418 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
420 static int __init find_node_by_addr(unsigned long addr)
422 int ret = NUMA_NO_NODE;
425 for_each_node_mask(i, nodes_parsed) {
427 * Find the real node that this emulated node appears on. For
428 * the sake of simplicity, we only use a real node's starting
429 * address to determine which emulated node it appears on.
431 if (addr >= nodes[i].start && addr < nodes[i].end) {
440 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
441 * mappings that respect the real ACPI topology but reflect our emulated
442 * environment. For each emulated node, we find which real node it appears on
443 * and create PXM to NID mappings for those fake nodes which mirror that
444 * locality. SLIT will now represent the correct distances between emulated
445 * nodes as a result of the real topology.
447 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
451 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
453 for (i = 0; i < num_nodes; i++) {
456 nid = find_node_by_addr(fake_nodes[i].start);
457 if (nid == NUMA_NO_NODE)
459 pxm = node_to_pxm(nid);
460 if (pxm == PXM_INVAL)
462 fake_node_to_pxm_map[i] = pxm;
464 * For each apicid_to_node mapping that exists for this real
465 * node, it must now point to the fake node ID.
467 for (j = 0; j < MAX_LOCAL_APIC; j++)
468 if (apicid_to_node[j] == nid)
469 fake_apicid_to_node[j] = i;
471 for (i = 0; i < num_nodes; i++)
472 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
473 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
475 nodes_clear(nodes_parsed);
476 for (i = 0; i < num_nodes; i++)
477 if (fake_nodes[i].start != fake_nodes[i].end)
478 node_set(i, nodes_parsed);
481 static int null_slit_node_compare(int a, int b)
483 return node_to_pxm(a) == node_to_pxm(b);
486 static int null_slit_node_compare(int a, int b)
490 #endif /* CONFIG_NUMA_EMU */
492 int __node_distance(int a, int b)
497 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
499 index = acpi_slit->locality_count * node_to_pxm(a);
500 return acpi_slit->entry[index + node_to_pxm(b)];
503 EXPORT_SYMBOL(__node_distance);
505 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
506 int memory_add_physaddr_to_nid(u64 start)
511 if (nodes_add[i].start <= start && nodes_add[i].end > start)
516 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);