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>
19 #include <linux/memblock.h>
21 #include <asm/proto.h>
25 #include <asm/uv/uv.h>
27 int acpi_numa __initdata;
29 static struct acpi_table_slit *acpi_slit;
31 static nodemask_t nodes_parsed __initdata;
32 static nodemask_t cpu_nodes_parsed __initdata;
33 static struct bootnode nodes[MAX_NUMNODES] __initdata;
34 static struct bootnode nodes_add[MAX_NUMNODES];
36 static int num_node_memblks __initdata;
37 static struct bootnode node_memblk_range[NR_NODE_MEMBLKS] __initdata;
38 static int memblk_nodeid[NR_NODE_MEMBLKS] __initdata;
40 static __init int setup_node(int pxm)
42 return acpi_map_pxm_to_node(pxm);
45 static __init int conflicting_memblks(unsigned long start, unsigned long end)
48 for (i = 0; i < num_node_memblks; i++) {
49 struct bootnode *nd = &node_memblk_range[i];
50 if (nd->start == nd->end)
52 if (nd->end > start && nd->start < end)
53 return memblk_nodeid[i];
54 if (nd->end == end && nd->start == start)
55 return memblk_nodeid[i];
60 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
62 struct bootnode *nd = &nodes[i];
64 if (nd->start < start) {
66 if (nd->end < nd->start)
71 if (nd->start > nd->end)
76 static __init void bad_srat(void)
79 printk(KERN_ERR "SRAT: SRAT not used.\n");
81 for (i = 0; i < MAX_LOCAL_APIC; i++)
82 apicid_to_node[i] = NUMA_NO_NODE;
83 for (i = 0; i < MAX_NUMNODES; i++) {
84 nodes[i].start = nodes[i].end = 0;
85 nodes_add[i].start = nodes_add[i].end = 0;
87 remove_all_active_ranges();
90 static __init inline int srat_disabled(void)
92 return numa_off || acpi_numa < 0;
95 /* Callback for SLIT parsing */
96 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
101 length = slit->header.length;
102 phys = memblock_find_in_range(0, max_pfn_mapped<<PAGE_SHIFT, length,
105 if (phys == MEMBLOCK_ERROR)
106 panic(" Can not save slit!\n");
108 acpi_slit = __va(phys);
109 memcpy(acpi_slit, slit, length);
110 memblock_x86_reserve_range(phys, phys + length, "ACPI SLIT");
113 /* Callback for Proximity Domain -> x2APIC mapping */
115 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
122 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
126 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
128 pxm = pa->proximity_domain;
129 node = setup_node(pxm);
131 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
136 apic_id = pa->apic_id;
137 apicid_to_node[apic_id] = node;
138 node_set(node, cpu_nodes_parsed);
140 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%04x -> Node %u\n",
144 /* Callback for Proximity Domain -> LAPIC mapping */
146 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
153 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
157 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
159 pxm = pa->proximity_domain_lo;
160 node = setup_node(pxm);
162 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
167 if (get_uv_system_type() >= UV_X2APIC)
168 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
170 apic_id = pa->apic_id;
171 apicid_to_node[apic_id] = node;
172 node_set(node, cpu_nodes_parsed);
174 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%02x -> Node %u\n",
178 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
179 static inline int save_add_info(void) {return 1;}
181 static inline int save_add_info(void) {return 0;}
185 * This code supports one contiguous hot add area per node
188 update_nodes_add(int node, unsigned long start, unsigned long end)
190 unsigned long s_pfn = start >> PAGE_SHIFT;
191 unsigned long e_pfn = end >> PAGE_SHIFT;
193 struct bootnode *nd = &nodes_add[node];
195 /* I had some trouble with strange memory hotadd regions breaking
196 the boot. Be very strict here and reject anything unexpected.
197 If you want working memory hotadd write correct SRATs.
199 The node size check is a basic sanity check to guard against
201 if ((signed long)(end - start) < NODE_MIN_SIZE) {
202 printk(KERN_ERR "SRAT: Hotplug area too small\n");
206 /* This check might be a bit too strict, but I'm keeping it for now. */
207 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
209 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
216 if (nd->start == nd->end) {
221 if (nd->start == end) {
225 if (nd->end == start) {
230 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
234 node_set(node, cpu_nodes_parsed);
235 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n",
240 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
242 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
244 struct bootnode *nd, oldnode;
245 unsigned long start, end;
251 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
255 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
258 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
260 start = ma->base_address;
261 end = start + ma->length;
262 pxm = ma->proximity_domain;
263 node = setup_node(pxm);
265 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
269 i = conflicting_memblks(start, end);
272 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
273 pxm, start, end, nodes[i].start, nodes[i].end);
276 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
277 pxm, start, end, node_to_pxm(i),
278 nodes[i].start, nodes[i].end);
284 if (!node_test_and_set(node, nodes_parsed)) {
288 if (start < nd->start)
294 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
297 if (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) {
298 update_nodes_add(node, start, end);
299 /* restore nodes[node] */
301 if ((nd->start | nd->end) == 0)
302 node_clear(node, nodes_parsed);
305 node_memblk_range[num_node_memblks].start = start;
306 node_memblk_range[num_node_memblks].end = end;
307 memblk_nodeid[num_node_memblks] = node;
311 /* Sanity check to catch more bad SRATs (they are amazingly common).
312 Make sure the PXMs cover all memory. */
313 static int __init nodes_cover_memory(const struct bootnode *nodes)
316 unsigned long pxmram, e820ram;
319 for_each_node_mask(i, nodes_parsed) {
320 unsigned long s = nodes[i].start >> PAGE_SHIFT;
321 unsigned long e = nodes[i].end >> PAGE_SHIFT;
323 pxmram -= __absent_pages_in_range(i, s, e);
324 if ((long)pxmram < 0)
328 e820ram = max_pfn - (memblock_x86_hole_size(0, max_pfn<<PAGE_SHIFT)>>PAGE_SHIFT);
329 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
330 if ((long)(e820ram - pxmram) >= (1<<(20 - PAGE_SHIFT))) {
332 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
333 (pxmram << PAGE_SHIFT) >> 20,
334 (e820ram << PAGE_SHIFT) >> 20);
340 void __init acpi_numa_arch_fixup(void) {}
342 int __init acpi_get_nodes(struct bootnode *physnodes)
347 for_each_node_mask(i, nodes_parsed) {
348 physnodes[ret].start = nodes[i].start;
349 physnodes[ret].end = nodes[i].end;
355 /* Use the information discovered above to actually set up the nodes. */
356 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
363 /* First clean up the node list */
364 for (i = 0; i < MAX_NUMNODES; i++)
365 cutoff_node(i, start, end);
368 * Join together blocks on the same node, holes between
369 * which don't overlap with memory on other nodes.
371 for (i = 0; i < num_node_memblks; ++i) {
374 for (j = i + 1; j < num_node_memblks; ++j) {
375 unsigned long start, end;
377 if (memblk_nodeid[i] != memblk_nodeid[j])
379 start = min(node_memblk_range[i].end,
380 node_memblk_range[j].end);
381 end = max(node_memblk_range[i].start,
382 node_memblk_range[j].start);
383 for (k = 0; k < num_node_memblks; ++k) {
384 if (memblk_nodeid[i] == memblk_nodeid[k])
386 if (start < node_memblk_range[k].end &&
387 end > node_memblk_range[k].start)
390 if (k < num_node_memblks)
392 start = min(node_memblk_range[i].start,
393 node_memblk_range[j].start);
394 end = max(node_memblk_range[i].end,
395 node_memblk_range[j].end);
396 printk(KERN_INFO "SRAT: Node %d "
397 "[%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n",
399 node_memblk_range[i].start,
400 node_memblk_range[i].end,
401 node_memblk_range[j].start,
402 node_memblk_range[j].end,
404 node_memblk_range[i].start = start;
405 node_memblk_range[i].end = end;
406 k = --num_node_memblks - j;
407 memmove(memblk_nodeid + j, memblk_nodeid + j+1,
408 k * sizeof(*memblk_nodeid));
409 memmove(node_memblk_range + j, node_memblk_range + j+1,
410 k * sizeof(*node_memblk_range));
415 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
417 if (memnode_shift < 0) {
419 "SRAT: No NUMA node hash function found. Contact maintainer\n");
424 for_each_node_mask(i, nodes_parsed)
425 memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
426 nodes[i].end >> PAGE_SHIFT);
427 /* for out of order entries in SRAT */
429 if (!nodes_cover_memory(nodes)) {
434 /* Account for nodes with cpus and no memory */
435 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed);
437 /* Finally register nodes */
438 for_each_node_mask(i, node_possible_map)
439 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
440 /* Try again in case setup_node_bootmem missed one due
441 to missing bootmem */
442 for_each_node_mask(i, node_possible_map)
444 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
446 for (i = 0; i < nr_cpu_ids; i++) {
447 int node = early_cpu_to_node(i);
449 if (node == NUMA_NO_NODE)
451 if (!node_online(node))
458 #ifdef CONFIG_NUMA_EMU
459 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
460 [0 ... MAX_NUMNODES-1] = PXM_INVAL
462 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
463 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
465 static int __init find_node_by_addr(unsigned long addr)
467 int ret = NUMA_NO_NODE;
470 for_each_node_mask(i, nodes_parsed) {
472 * Find the real node that this emulated node appears on. For
473 * the sake of simplicity, we only use a real node's starting
474 * address to determine which emulated node it appears on.
476 if (addr >= nodes[i].start && addr < nodes[i].end) {
485 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
486 * mappings that respect the real ACPI topology but reflect our emulated
487 * environment. For each emulated node, we find which real node it appears on
488 * and create PXM to NID mappings for those fake nodes which mirror that
489 * locality. SLIT will now represent the correct distances between emulated
490 * nodes as a result of the real topology.
492 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
496 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
498 for (i = 0; i < num_nodes; i++) {
501 nid = find_node_by_addr(fake_nodes[i].start);
502 if (nid == NUMA_NO_NODE)
504 pxm = node_to_pxm(nid);
505 if (pxm == PXM_INVAL)
507 fake_node_to_pxm_map[i] = pxm;
509 * For each apicid_to_node mapping that exists for this real
510 * node, it must now point to the fake node ID.
512 for (j = 0; j < MAX_LOCAL_APIC; j++)
513 if (apicid_to_node[j] == nid &&
514 fake_apicid_to_node[j] == NUMA_NO_NODE)
515 fake_apicid_to_node[j] = i;
517 for (i = 0; i < num_nodes; i++)
518 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
519 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
521 nodes_clear(nodes_parsed);
522 for (i = 0; i < num_nodes; i++)
523 if (fake_nodes[i].start != fake_nodes[i].end)
524 node_set(i, nodes_parsed);
527 static int null_slit_node_compare(int a, int b)
529 return node_to_pxm(a) == node_to_pxm(b);
532 static int null_slit_node_compare(int a, int b)
536 #endif /* CONFIG_NUMA_EMU */
538 int __node_distance(int a, int b)
543 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
545 index = acpi_slit->locality_count * node_to_pxm(a);
546 return acpi_slit->entry[index + node_to_pxm(b)];
549 EXPORT_SYMBOL(__node_distance);
551 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
552 int memory_add_physaddr_to_nid(u64 start)
557 if (nodes_add[i].start <= start && nodes_add[i].end > start)
562 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);