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>
25 int acpi_numa __initdata;
27 static struct acpi_table_slit *acpi_slit;
29 static nodemask_t nodes_parsed __initdata;
30 static struct bootnode nodes[MAX_NUMNODES] __initdata;
31 static struct bootnode nodes_add[MAX_NUMNODES];
32 static int found_add_area __initdata;
33 int hotadd_percent __initdata = 0;
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 /* Too small nodes confuse the VM badly. Usually they result
41 #define NODE_MIN_SIZE (4*1024*1024)
43 static __init int setup_node(int pxm)
45 return acpi_map_pxm_to_node(pxm);
48 static __init int conflicting_memblks(unsigned long start, unsigned long end)
51 for (i = 0; i < num_node_memblks; i++) {
52 struct bootnode *nd = &node_memblk_range[i];
53 if (nd->start == nd->end)
55 if (nd->end > start && nd->start < end)
56 return memblk_nodeid[i];
57 if (nd->end == end && nd->start == start)
58 return memblk_nodeid[i];
63 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
65 struct bootnode *nd = &nodes[i];
70 if (nd->start < start) {
72 if (nd->end < nd->start)
77 if (nd->start > nd->end)
82 static __init void bad_srat(void)
85 printk(KERN_ERR "SRAT: SRAT not used.\n");
88 for (i = 0; i < MAX_LOCAL_APIC; i++)
89 apicid_to_node[i] = NUMA_NO_NODE;
90 for (i = 0; i < MAX_NUMNODES; i++)
91 nodes_add[i].start = nodes[i].end = 0;
92 remove_all_active_ranges();
95 static __init inline int srat_disabled(void)
97 return numa_off || acpi_numa < 0;
100 /* Callback for SLIT parsing */
101 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
106 length = slit->header.length;
107 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length,
111 panic(" Can not save slit!\n");
113 acpi_slit = __va(phys);
114 memcpy(acpi_slit, slit, length);
115 reserve_early(phys, phys + length, "ACPI SLIT");
118 /* Callback for Proximity Domain -> LAPIC mapping */
120 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
127 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
131 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
133 pxm = pa->proximity_domain_lo;
134 node = setup_node(pxm);
136 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
142 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
144 apic_id = pa->apic_id;
145 apicid_to_node[apic_id] = node;
147 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
151 static int update_end_of_memory(unsigned long end) {return -1;}
152 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
153 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
154 static inline int save_add_info(void) {return 1;}
156 static inline int save_add_info(void) {return 0;}
159 * Update nodes_add and decide if to include add are in the zone.
160 * Both SPARSE and RESERVE need nodes_add information.
161 * This code supports one contiguous hot add area per node.
164 reserve_hotadd(int node, unsigned long start, unsigned long end)
166 unsigned long s_pfn = start >> PAGE_SHIFT;
167 unsigned long e_pfn = end >> PAGE_SHIFT;
168 int ret = 0, changed = 0;
169 struct bootnode *nd = &nodes_add[node];
171 /* I had some trouble with strange memory hotadd regions breaking
172 the boot. Be very strict here and reject anything unexpected.
173 If you want working memory hotadd write correct SRATs.
175 The node size check is a basic sanity check to guard against
177 if ((signed long)(end - start) < NODE_MIN_SIZE) {
178 printk(KERN_ERR "SRAT: Hotplug area too small\n");
182 /* This check might be a bit too strict, but I'm keeping it for now. */
183 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
185 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
190 if (!hotadd_enough_memory(&nodes_add[node])) {
191 printk(KERN_ERR "SRAT: Hotplug area too large\n");
197 if (nd->start == nd->end) {
202 if (nd->start == end) {
206 if (nd->end == start) {
211 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
214 ret = update_end_of_memory(nd->end);
217 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
221 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
223 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
225 struct bootnode *nd, oldnode;
226 unsigned long start, end;
232 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
236 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
239 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
241 start = ma->base_address;
242 end = start + ma->length;
243 pxm = ma->proximity_domain;
244 node = setup_node(pxm);
246 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
250 i = conflicting_memblks(start, end);
253 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
254 pxm, start, end, nodes[i].start, nodes[i].end);
257 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
258 pxm, start, end, node_to_pxm(i),
259 nodes[i].start, nodes[i].end);
265 if (!node_test_and_set(node, nodes_parsed)) {
269 if (start < nd->start)
275 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
277 e820_register_active_regions(node, start >> PAGE_SHIFT,
279 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
280 nd->end >> PAGE_SHIFT);
282 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
283 (reserve_hotadd(node, start, end) < 0)) {
284 /* Ignore hotadd region. Undo damage */
285 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
287 if ((nd->start | nd->end) == 0)
288 node_clear(node, nodes_parsed);
291 node_memblk_range[num_node_memblks].start = start;
292 node_memblk_range[num_node_memblks].end = end;
293 memblk_nodeid[num_node_memblks] = node;
297 /* Sanity check to catch more bad SRATs (they are amazingly common).
298 Make sure the PXMs cover all memory. */
299 static int __init nodes_cover_memory(const struct bootnode *nodes)
302 unsigned long pxmram, e820ram;
305 for_each_node_mask(i, nodes_parsed) {
306 unsigned long s = nodes[i].start >> PAGE_SHIFT;
307 unsigned long e = nodes[i].end >> PAGE_SHIFT;
309 pxmram -= absent_pages_in_range(s, e);
310 if ((long)pxmram < 0)
314 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
315 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
316 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
318 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
319 (pxmram << PAGE_SHIFT) >> 20,
320 (e820ram << PAGE_SHIFT) >> 20);
326 static void __init unparse_node(int node)
329 node_clear(node, nodes_parsed);
330 for (i = 0; i < MAX_LOCAL_APIC; i++) {
331 if (apicid_to_node[i] == node)
332 apicid_to_node[i] = NUMA_NO_NODE;
336 void __init acpi_numa_arch_fixup(void) {}
338 /* Use the information discovered above to actually set up the nodes. */
339 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
346 /* First clean up the node list */
347 for (i = 0; i < MAX_NUMNODES; i++) {
348 cutoff_node(i, start, end);
350 * don't confuse VM with a node that doesn't have the
354 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
360 if (!nodes_cover_memory(nodes)) {
365 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
367 if (memnode_shift < 0) {
369 "SRAT: No NUMA node hash function found. Contact maintainer\n");
374 node_possible_map = nodes_parsed;
376 /* Finally register nodes */
377 for_each_node_mask(i, node_possible_map)
378 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
379 /* Try again in case setup_node_bootmem missed one due
380 to missing bootmem */
381 for_each_node_mask(i, node_possible_map)
383 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
385 for (i = 0; i < NR_CPUS; i++) {
386 int node = early_cpu_to_node(i);
388 if (node == NUMA_NO_NODE)
390 if (!node_isset(node, node_possible_map))
397 #ifdef CONFIG_NUMA_EMU
398 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
399 [0 ... MAX_NUMNODES-1] = PXM_INVAL
401 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
402 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
404 static int __init find_node_by_addr(unsigned long addr)
406 int ret = NUMA_NO_NODE;
409 for_each_node_mask(i, nodes_parsed) {
411 * Find the real node that this emulated node appears on. For
412 * the sake of simplicity, we only use a real node's starting
413 * address to determine which emulated node it appears on.
415 if (addr >= nodes[i].start && addr < nodes[i].end) {
424 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
425 * mappings that respect the real ACPI topology but reflect our emulated
426 * environment. For each emulated node, we find which real node it appears on
427 * and create PXM to NID mappings for those fake nodes which mirror that
428 * locality. SLIT will now represent the correct distances between emulated
429 * nodes as a result of the real topology.
431 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
435 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
437 for (i = 0; i < num_nodes; i++) {
440 nid = find_node_by_addr(fake_nodes[i].start);
441 if (nid == NUMA_NO_NODE)
443 pxm = node_to_pxm(nid);
444 if (pxm == PXM_INVAL)
446 fake_node_to_pxm_map[i] = pxm;
448 * For each apicid_to_node mapping that exists for this real
449 * node, it must now point to the fake node ID.
451 for (j = 0; j < MAX_LOCAL_APIC; j++)
452 if (apicid_to_node[j] == nid)
453 fake_apicid_to_node[j] = i;
455 for (i = 0; i < num_nodes; i++)
456 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
457 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
459 nodes_clear(nodes_parsed);
460 for (i = 0; i < num_nodes; i++)
461 if (fake_nodes[i].start != fake_nodes[i].end)
462 node_set(i, nodes_parsed);
463 WARN_ON(!nodes_cover_memory(fake_nodes));
466 static int null_slit_node_compare(int a, int b)
468 return node_to_pxm(a) == node_to_pxm(b);
471 static int null_slit_node_compare(int a, int b)
475 #endif /* CONFIG_NUMA_EMU */
477 void __init srat_reserve_add_area(int nodeid)
479 if (found_add_area && nodes_add[nodeid].end) {
482 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
483 "for node %d at %Lx-%Lx\n",
484 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
485 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
487 total_mb *= sizeof(struct page);
489 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
490 "pre-allocated memory.\n", (unsigned long long)total_mb);
491 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
492 nodes_add[nodeid].end - nodes_add[nodeid].start,
497 int __node_distance(int a, int b)
502 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
504 index = acpi_slit->locality_count * node_to_pxm(a);
505 return acpi_slit->entry[index + node_to_pxm(b)];
508 EXPORT_SYMBOL(__node_distance);
510 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
511 int memory_add_physaddr_to_nid(u64 start)
516 if (nodes_add[i].start <= start && nodes_add[i].end > start)
521 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);