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>
23 #include <asm/genapic.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 struct bootnode nodes[MAX_NUMNODES] __initdata;
32 static struct bootnode nodes_add[MAX_NUMNODES];
33 static int found_add_area __initdata;
34 int hotadd_percent __initdata = 0;
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 /* Too small nodes confuse the VM badly. Usually they result
42 #define NODE_MIN_SIZE (4*1024*1024)
44 static __init int setup_node(int pxm)
46 return acpi_map_pxm_to_node(pxm);
49 static __init int conflicting_memblks(unsigned long start, unsigned long end)
52 for (i = 0; i < num_node_memblks; i++) {
53 struct bootnode *nd = &node_memblk_range[i];
54 if (nd->start == nd->end)
56 if (nd->end > start && nd->start < end)
57 return memblk_nodeid[i];
58 if (nd->end == end && nd->start == start)
59 return memblk_nodeid[i];
64 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
66 struct bootnode *nd = &nodes[i];
71 if (nd->start < start) {
73 if (nd->end < nd->start)
78 if (nd->start > nd->end)
83 static __init void bad_srat(void)
86 printk(KERN_ERR "SRAT: SRAT not used.\n");
89 for (i = 0; i < MAX_LOCAL_APIC; i++)
90 apicid_to_node[i] = NUMA_NO_NODE;
91 for (i = 0; i < MAX_NUMNODES; i++)
92 nodes_add[i].start = nodes[i].end = 0;
93 remove_all_active_ranges();
96 static __init inline int srat_disabled(void)
98 return numa_off || acpi_numa < 0;
101 /* Callback for SLIT parsing */
102 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
107 length = slit->header.length;
108 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
112 panic(" Can not save slit!\n");
114 acpi_slit = __va(phys);
115 memcpy(acpi_slit, slit, length);
116 reserve_early(phys, phys + length, "ACPI SLIT");
119 /* Callback for Proximity Domain -> LAPIC mapping */
121 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
128 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
132 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
134 pxm = pa->proximity_domain_lo;
135 node = setup_node(pxm);
137 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
142 if (get_uv_system_type() >= UV_X2APIC)
143 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
145 apic_id = pa->apic_id;
146 apicid_to_node[apic_id] = node;
148 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
152 static int update_end_of_memory(unsigned long end) {return -1;}
153 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
154 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
155 static inline int save_add_info(void) {return 1;}
157 static inline int save_add_info(void) {return 0;}
160 * Update nodes_add and decide if to include add are in the zone.
161 * Both SPARSE and RESERVE need nodes_add information.
162 * This code supports one contiguous hot add area per node.
165 reserve_hotadd(int node, unsigned long start, unsigned long end)
167 unsigned long s_pfn = start >> PAGE_SHIFT;
168 unsigned long e_pfn = end >> PAGE_SHIFT;
169 int ret = 0, changed = 0;
170 struct bootnode *nd = &nodes_add[node];
172 /* I had some trouble with strange memory hotadd regions breaking
173 the boot. Be very strict here and reject anything unexpected.
174 If you want working memory hotadd write correct SRATs.
176 The node size check is a basic sanity check to guard against
178 if ((signed long)(end - start) < NODE_MIN_SIZE) {
179 printk(KERN_ERR "SRAT: Hotplug area too small\n");
183 /* This check might be a bit too strict, but I'm keeping it for now. */
184 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
186 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
191 if (!hotadd_enough_memory(&nodes_add[node])) {
192 printk(KERN_ERR "SRAT: Hotplug area too large\n");
198 if (nd->start == nd->end) {
203 if (nd->start == end) {
207 if (nd->end == start) {
212 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
215 ret = update_end_of_memory(nd->end);
218 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
222 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
224 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
226 struct bootnode *nd, oldnode;
227 unsigned long start, end;
233 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
237 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
240 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
242 start = ma->base_address;
243 end = start + ma->length;
244 pxm = ma->proximity_domain;
245 node = setup_node(pxm);
247 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
251 i = conflicting_memblks(start, end);
254 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
255 pxm, start, end, nodes[i].start, nodes[i].end);
258 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
259 pxm, start, end, node_to_pxm(i),
260 nodes[i].start, nodes[i].end);
266 if (!node_test_and_set(node, nodes_parsed)) {
270 if (start < nd->start)
276 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
278 e820_register_active_regions(node, start >> PAGE_SHIFT,
280 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
281 nd->end >> PAGE_SHIFT);
283 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
284 (reserve_hotadd(node, start, end) < 0)) {
285 /* Ignore hotadd region. Undo damage */
286 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
288 if ((nd->start | nd->end) == 0)
289 node_clear(node, nodes_parsed);
292 node_memblk_range[num_node_memblks].start = start;
293 node_memblk_range[num_node_memblks].end = end;
294 memblk_nodeid[num_node_memblks] = node;
298 /* Sanity check to catch more bad SRATs (they are amazingly common).
299 Make sure the PXMs cover all memory. */
300 static int __init nodes_cover_memory(const struct bootnode *nodes)
303 unsigned long pxmram, e820ram;
306 for_each_node_mask(i, nodes_parsed) {
307 unsigned long s = nodes[i].start >> PAGE_SHIFT;
308 unsigned long e = nodes[i].end >> PAGE_SHIFT;
310 pxmram -= absent_pages_in_range(s, e);
311 if ((long)pxmram < 0)
315 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
316 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
317 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
319 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
320 (pxmram << PAGE_SHIFT) >> 20,
321 (e820ram << PAGE_SHIFT) >> 20);
327 static void __init unparse_node(int node)
330 node_clear(node, nodes_parsed);
331 for (i = 0; i < MAX_LOCAL_APIC; i++) {
332 if (apicid_to_node[i] == node)
333 apicid_to_node[i] = NUMA_NO_NODE;
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 * don't confuse VM with a node that doesn't have the
355 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
361 if (!nodes_cover_memory(nodes)) {
366 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
368 if (memnode_shift < 0) {
370 "SRAT: No NUMA node hash function found. Contact maintainer\n");
375 node_possible_map = nodes_parsed;
377 /* Finally register nodes */
378 for_each_node_mask(i, node_possible_map)
379 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
380 /* Try again in case setup_node_bootmem missed one due
381 to missing bootmem */
382 for_each_node_mask(i, node_possible_map)
384 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
386 for (i = 0; i < nr_cpu_ids; i++) {
387 int node = early_cpu_to_node(i);
389 if (node == NUMA_NO_NODE)
391 if (!node_isset(node, node_possible_map))
398 #ifdef CONFIG_NUMA_EMU
399 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
400 [0 ... MAX_NUMNODES-1] = PXM_INVAL
402 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
403 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
405 static int __init find_node_by_addr(unsigned long addr)
407 int ret = NUMA_NO_NODE;
410 for_each_node_mask(i, nodes_parsed) {
412 * Find the real node that this emulated node appears on. For
413 * the sake of simplicity, we only use a real node's starting
414 * address to determine which emulated node it appears on.
416 if (addr >= nodes[i].start && addr < nodes[i].end) {
425 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
426 * mappings that respect the real ACPI topology but reflect our emulated
427 * environment. For each emulated node, we find which real node it appears on
428 * and create PXM to NID mappings for those fake nodes which mirror that
429 * locality. SLIT will now represent the correct distances between emulated
430 * nodes as a result of the real topology.
432 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
436 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
438 for (i = 0; i < num_nodes; i++) {
441 nid = find_node_by_addr(fake_nodes[i].start);
442 if (nid == NUMA_NO_NODE)
444 pxm = node_to_pxm(nid);
445 if (pxm == PXM_INVAL)
447 fake_node_to_pxm_map[i] = pxm;
449 * For each apicid_to_node mapping that exists for this real
450 * node, it must now point to the fake node ID.
452 for (j = 0; j < MAX_LOCAL_APIC; j++)
453 if (apicid_to_node[j] == nid)
454 fake_apicid_to_node[j] = i;
456 for (i = 0; i < num_nodes; i++)
457 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
458 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
460 nodes_clear(nodes_parsed);
461 for (i = 0; i < num_nodes; i++)
462 if (fake_nodes[i].start != fake_nodes[i].end)
463 node_set(i, nodes_parsed);
464 WARN_ON(!nodes_cover_memory(fake_nodes));
467 static int null_slit_node_compare(int a, int b)
469 return node_to_pxm(a) == node_to_pxm(b);
472 static int null_slit_node_compare(int a, int b)
476 #endif /* CONFIG_NUMA_EMU */
478 void __init srat_reserve_add_area(int nodeid)
480 if (found_add_area && nodes_add[nodeid].end) {
483 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
484 "for node %d at %Lx-%Lx\n",
485 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
486 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
488 total_mb *= sizeof(struct page);
490 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
491 "pre-allocated memory.\n", (unsigned long long)total_mb);
492 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
493 nodes_add[nodeid].end - nodes_add[nodeid].start,
498 int __node_distance(int a, int b)
503 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
505 index = acpi_slit->locality_count * node_to_pxm(a);
506 return acpi_slit->entry[index + node_to_pxm(b)];
509 EXPORT_SYMBOL(__node_distance);
511 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
512 int memory_add_physaddr_to_nid(u64 start)
517 if (nodes_add[i].start <= start && nodes_add[i].end > start)
522 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);