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];
34 static int found_add_area __initdata;
35 int hotadd_percent __initdata = 0;
37 static int num_node_memblks __initdata;
38 static struct bootnode node_memblk_range[NR_NODE_MEMBLKS] __initdata;
39 static int memblk_nodeid[NR_NODE_MEMBLKS] __initdata;
41 /* Too small nodes confuse the VM badly. Usually they result
43 #define NODE_MIN_SIZE (4*1024*1024)
45 static __init int setup_node(int pxm)
47 return acpi_map_pxm_to_node(pxm);
50 static __init int conflicting_memblks(unsigned long start, unsigned long end)
53 for (i = 0; i < num_node_memblks; i++) {
54 struct bootnode *nd = &node_memblk_range[i];
55 if (nd->start == nd->end)
57 if (nd->end > start && nd->start < end)
58 return memblk_nodeid[i];
59 if (nd->end == end && nd->start == start)
60 return memblk_nodeid[i];
65 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
67 struct bootnode *nd = &nodes[i];
72 if (nd->start < start) {
74 if (nd->end < nd->start)
79 if (nd->start > nd->end)
84 static __init void bad_srat(void)
87 printk(KERN_ERR "SRAT: SRAT not used.\n");
90 for (i = 0; i < MAX_LOCAL_APIC; i++)
91 apicid_to_node[i] = NUMA_NO_NODE;
92 for (i = 0; i < MAX_NUMNODES; i++)
93 nodes_add[i].start = nodes[i].end = 0;
94 remove_all_active_ranges();
97 static __init inline int srat_disabled(void)
99 return numa_off || acpi_numa < 0;
102 /* Callback for SLIT parsing */
103 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
108 length = slit->header.length;
109 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
113 panic(" Can not save slit!\n");
115 acpi_slit = __va(phys);
116 memcpy(acpi_slit, slit, length);
117 reserve_early(phys, phys + length, "ACPI SLIT");
120 /* Callback for Proximity Domain -> x2APIC mapping */
122 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
129 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
133 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
135 pxm = pa->proximity_domain;
136 node = setup_node(pxm);
138 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
143 apic_id = pa->apic_id;
144 apicid_to_node[apic_id] = node;
145 node_set(node, cpu_nodes_parsed);
147 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
151 /* Callback for Proximity Domain -> LAPIC mapping */
153 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
160 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
164 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
166 pxm = pa->proximity_domain_lo;
167 node = setup_node(pxm);
169 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
174 if (get_uv_system_type() >= UV_X2APIC)
175 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
177 apic_id = pa->apic_id;
178 apicid_to_node[apic_id] = node;
179 node_set(node, cpu_nodes_parsed);
181 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
185 static int update_end_of_memory(unsigned long end) {return -1;}
186 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
187 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
188 static inline int save_add_info(void) {return 1;}
190 static inline int save_add_info(void) {return 0;}
193 * Update nodes_add and decide if to include add are in the zone.
194 * Both SPARSE and RESERVE need nodes_add information.
195 * This code supports one contiguous hot add area per node.
198 reserve_hotadd(int node, unsigned long start, unsigned long end)
200 unsigned long s_pfn = start >> PAGE_SHIFT;
201 unsigned long e_pfn = end >> PAGE_SHIFT;
202 int ret = 0, changed = 0;
203 struct bootnode *nd = &nodes_add[node];
205 /* I had some trouble with strange memory hotadd regions breaking
206 the boot. Be very strict here and reject anything unexpected.
207 If you want working memory hotadd write correct SRATs.
209 The node size check is a basic sanity check to guard against
211 if ((signed long)(end - start) < NODE_MIN_SIZE) {
212 printk(KERN_ERR "SRAT: Hotplug area too small\n");
216 /* This check might be a bit too strict, but I'm keeping it for now. */
217 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
219 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
224 if (!hotadd_enough_memory(&nodes_add[node])) {
225 printk(KERN_ERR "SRAT: Hotplug area too large\n");
231 if (nd->start == nd->end) {
236 if (nd->start == end) {
240 if (nd->end == start) {
245 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
248 ret = update_end_of_memory(nd->end);
251 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
255 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
257 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
259 struct bootnode *nd, oldnode;
260 unsigned long start, end;
266 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
270 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
273 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
275 start = ma->base_address;
276 end = start + ma->length;
277 pxm = ma->proximity_domain;
278 node = setup_node(pxm);
280 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
284 i = conflicting_memblks(start, end);
287 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
288 pxm, start, end, nodes[i].start, nodes[i].end);
291 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
292 pxm, start, end, node_to_pxm(i),
293 nodes[i].start, nodes[i].end);
299 if (!node_test_and_set(node, nodes_parsed)) {
303 if (start < nd->start)
309 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
311 e820_register_active_regions(node, start >> PAGE_SHIFT,
313 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
314 nd->end >> PAGE_SHIFT);
316 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
317 (reserve_hotadd(node, start, end) < 0)) {
318 /* Ignore hotadd region. Undo damage */
319 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
321 if ((nd->start | nd->end) == 0)
322 node_clear(node, nodes_parsed);
325 node_memblk_range[num_node_memblks].start = start;
326 node_memblk_range[num_node_memblks].end = end;
327 memblk_nodeid[num_node_memblks] = node;
331 /* Sanity check to catch more bad SRATs (they are amazingly common).
332 Make sure the PXMs cover all memory. */
333 static int __init nodes_cover_memory(const struct bootnode *nodes)
336 unsigned long pxmram, e820ram;
339 for_each_node_mask(i, nodes_parsed) {
340 unsigned long s = nodes[i].start >> PAGE_SHIFT;
341 unsigned long e = nodes[i].end >> PAGE_SHIFT;
343 pxmram -= absent_pages_in_range(s, e);
344 if ((long)pxmram < 0)
348 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
349 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
350 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
352 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
353 (pxmram << PAGE_SHIFT) >> 20,
354 (e820ram << PAGE_SHIFT) >> 20);
360 static void __init unparse_node(int node)
363 node_clear(node, nodes_parsed);
364 node_clear(node, cpu_nodes_parsed);
365 for (i = 0; i < MAX_LOCAL_APIC; i++) {
366 if (apicid_to_node[i] == node)
367 apicid_to_node[i] = NUMA_NO_NODE;
371 void __init acpi_numa_arch_fixup(void) {}
373 /* Use the information discovered above to actually set up the nodes. */
374 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
381 /* First clean up the node list */
382 for (i = 0; i < MAX_NUMNODES; i++) {
383 cutoff_node(i, start, end);
385 * don't confuse VM with a node that doesn't have the
389 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
395 if (!nodes_cover_memory(nodes)) {
400 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
402 if (memnode_shift < 0) {
404 "SRAT: No NUMA node hash function found. Contact maintainer\n");
409 /* Account for nodes with cpus and no memory */
410 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed);
412 /* Finally register nodes */
413 for_each_node_mask(i, node_possible_map)
414 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
415 /* Try again in case setup_node_bootmem missed one due
416 to missing bootmem */
417 for_each_node_mask(i, node_possible_map)
419 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
421 for (i = 0; i < nr_cpu_ids; i++) {
422 int node = early_cpu_to_node(i);
424 if (node == NUMA_NO_NODE)
426 if (!node_isset(node, node_possible_map))
433 #ifdef CONFIG_NUMA_EMU
434 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
435 [0 ... MAX_NUMNODES-1] = PXM_INVAL
437 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
438 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
440 static int __init find_node_by_addr(unsigned long addr)
442 int ret = NUMA_NO_NODE;
445 for_each_node_mask(i, nodes_parsed) {
447 * Find the real node that this emulated node appears on. For
448 * the sake of simplicity, we only use a real node's starting
449 * address to determine which emulated node it appears on.
451 if (addr >= nodes[i].start && addr < nodes[i].end) {
460 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
461 * mappings that respect the real ACPI topology but reflect our emulated
462 * environment. For each emulated node, we find which real node it appears on
463 * and create PXM to NID mappings for those fake nodes which mirror that
464 * locality. SLIT will now represent the correct distances between emulated
465 * nodes as a result of the real topology.
467 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
471 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
473 for (i = 0; i < num_nodes; i++) {
476 nid = find_node_by_addr(fake_nodes[i].start);
477 if (nid == NUMA_NO_NODE)
479 pxm = node_to_pxm(nid);
480 if (pxm == PXM_INVAL)
482 fake_node_to_pxm_map[i] = pxm;
484 * For each apicid_to_node mapping that exists for this real
485 * node, it must now point to the fake node ID.
487 for (j = 0; j < MAX_LOCAL_APIC; j++)
488 if (apicid_to_node[j] == nid)
489 fake_apicid_to_node[j] = i;
491 for (i = 0; i < num_nodes; i++)
492 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
493 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
495 nodes_clear(nodes_parsed);
496 for (i = 0; i < num_nodes; i++)
497 if (fake_nodes[i].start != fake_nodes[i].end)
498 node_set(i, nodes_parsed);
499 WARN_ON(!nodes_cover_memory(fake_nodes));
502 static int null_slit_node_compare(int a, int b)
504 return node_to_pxm(a) == node_to_pxm(b);
507 static int null_slit_node_compare(int a, int b)
511 #endif /* CONFIG_NUMA_EMU */
513 void __init srat_reserve_add_area(int nodeid)
515 if (found_add_area && nodes_add[nodeid].end) {
518 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
519 "for node %d at %Lx-%Lx\n",
520 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
521 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
523 total_mb *= sizeof(struct page);
525 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
526 "pre-allocated memory.\n", (unsigned long long)total_mb);
527 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
528 nodes_add[nodeid].end - nodes_add[nodeid].start,
533 int __node_distance(int a, int b)
538 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
540 index = acpi_slit->locality_count * node_to_pxm(a);
541 return acpi_slit->entry[index + node_to_pxm(b)];
544 EXPORT_SYMBOL(__node_distance);
546 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
547 int memory_add_physaddr_to_nid(u64 start)
552 if (nodes_add[i].start <= start && nodes_add[i].end > start)
557 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);