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[i].start = nodes[i].end = 0;
92 nodes_add[i].start = nodes_add[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 -> LAPIC mapping */
122 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
129 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
133 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
135 pxm = pa->proximity_domain_lo;
136 node = setup_node(pxm);
138 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
144 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
146 apic_id = pa->apic_id;
147 apicid_to_node[apic_id] = node;
149 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
153 static int update_end_of_memory(unsigned long end) {return -1;}
154 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
155 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
156 static inline int save_add_info(void) {return 1;}
158 static inline int save_add_info(void) {return 0;}
161 * Update nodes_add and decide if to include add are in the zone.
162 * Both SPARSE and RESERVE need nodes_add information.
163 * This code supports one contiguous hot add area per node.
166 reserve_hotadd(int node, unsigned long start, unsigned long end)
168 unsigned long s_pfn = start >> PAGE_SHIFT;
169 unsigned long e_pfn = end >> PAGE_SHIFT;
170 int ret = 0, changed = 0;
171 struct bootnode *nd = &nodes_add[node];
173 /* I had some trouble with strange memory hotadd regions breaking
174 the boot. Be very strict here and reject anything unexpected.
175 If you want working memory hotadd write correct SRATs.
177 The node size check is a basic sanity check to guard against
179 if ((signed long)(end - start) < NODE_MIN_SIZE) {
180 printk(KERN_ERR "SRAT: Hotplug area too small\n");
184 /* This check might be a bit too strict, but I'm keeping it for now. */
185 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
187 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
192 if (!hotadd_enough_memory(&nodes_add[node])) {
193 printk(KERN_ERR "SRAT: Hotplug area too large\n");
199 if (nd->start == nd->end) {
204 if (nd->start == end) {
208 if (nd->end == start) {
213 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
216 ret = update_end_of_memory(nd->end);
219 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
223 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
225 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
227 struct bootnode *nd, oldnode;
228 unsigned long start, end;
234 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
238 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
241 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
243 start = ma->base_address;
244 end = start + ma->length;
245 pxm = ma->proximity_domain;
246 node = setup_node(pxm);
248 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
252 i = conflicting_memblks(start, end);
255 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
256 pxm, start, end, nodes[i].start, nodes[i].end);
259 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
260 pxm, start, end, node_to_pxm(i),
261 nodes[i].start, nodes[i].end);
267 if (!node_test_and_set(node, nodes_parsed)) {
271 if (start < nd->start)
277 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
279 e820_register_active_regions(node, start >> PAGE_SHIFT,
281 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
282 nd->end >> PAGE_SHIFT);
284 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
285 (reserve_hotadd(node, start, end) < 0)) {
286 /* Ignore hotadd region. Undo damage */
287 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
289 if ((nd->start | nd->end) == 0)
290 node_clear(node, nodes_parsed);
293 node_memblk_range[num_node_memblks].start = start;
294 node_memblk_range[num_node_memblks].end = end;
295 memblk_nodeid[num_node_memblks] = node;
299 /* Sanity check to catch more bad SRATs (they are amazingly common).
300 Make sure the PXMs cover all memory. */
301 static int __init nodes_cover_memory(const struct bootnode *nodes)
304 unsigned long pxmram, e820ram;
307 for_each_node_mask(i, nodes_parsed) {
308 unsigned long s = nodes[i].start >> PAGE_SHIFT;
309 unsigned long e = nodes[i].end >> PAGE_SHIFT;
311 pxmram -= absent_pages_in_range(s, e);
312 if ((long)pxmram < 0)
316 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
317 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
318 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
320 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
321 (pxmram << PAGE_SHIFT) >> 20,
322 (e820ram << PAGE_SHIFT) >> 20);
328 static void __init unparse_node(int node)
331 node_clear(node, nodes_parsed);
332 for (i = 0; i < MAX_LOCAL_APIC; i++) {
333 if (apicid_to_node[i] == node)
334 apicid_to_node[i] = NUMA_NO_NODE;
338 void __init acpi_numa_arch_fixup(void) {}
340 /* Use the information discovered above to actually set up the nodes. */
341 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
348 /* First clean up the node list */
349 for (i = 0; i < MAX_NUMNODES; i++) {
350 cutoff_node(i, start, end);
352 * don't confuse VM with a node that doesn't have the
356 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
362 if (!nodes_cover_memory(nodes)) {
367 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
369 if (memnode_shift < 0) {
371 "SRAT: No NUMA node hash function found. Contact maintainer\n");
376 node_possible_map = nodes_parsed;
378 /* Finally register nodes */
379 for_each_node_mask(i, node_possible_map)
380 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
381 /* Try again in case setup_node_bootmem missed one due
382 to missing bootmem */
383 for_each_node_mask(i, node_possible_map)
385 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
387 for (i = 0; i < NR_CPUS; i++) {
388 int node = early_cpu_to_node(i);
390 if (node == NUMA_NO_NODE)
392 if (!node_isset(node, node_possible_map))
399 #ifdef CONFIG_NUMA_EMU
400 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
401 [0 ... MAX_NUMNODES-1] = PXM_INVAL
403 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
404 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
406 static int __init find_node_by_addr(unsigned long addr)
408 int ret = NUMA_NO_NODE;
411 for_each_node_mask(i, nodes_parsed) {
413 * Find the real node that this emulated node appears on. For
414 * the sake of simplicity, we only use a real node's starting
415 * address to determine which emulated node it appears on.
417 if (addr >= nodes[i].start && addr < nodes[i].end) {
426 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
427 * mappings that respect the real ACPI topology but reflect our emulated
428 * environment. For each emulated node, we find which real node it appears on
429 * and create PXM to NID mappings for those fake nodes which mirror that
430 * locality. SLIT will now represent the correct distances between emulated
431 * nodes as a result of the real topology.
433 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
437 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
439 for (i = 0; i < num_nodes; i++) {
442 nid = find_node_by_addr(fake_nodes[i].start);
443 if (nid == NUMA_NO_NODE)
445 pxm = node_to_pxm(nid);
446 if (pxm == PXM_INVAL)
448 fake_node_to_pxm_map[i] = pxm;
450 * For each apicid_to_node mapping that exists for this real
451 * node, it must now point to the fake node ID.
453 for (j = 0; j < MAX_LOCAL_APIC; j++)
454 if (apicid_to_node[j] == nid)
455 fake_apicid_to_node[j] = i;
457 for (i = 0; i < num_nodes; i++)
458 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
459 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
461 nodes_clear(nodes_parsed);
462 for (i = 0; i < num_nodes; i++)
463 if (fake_nodes[i].start != fake_nodes[i].end)
464 node_set(i, nodes_parsed);
465 WARN_ON(!nodes_cover_memory(fake_nodes));
468 static int null_slit_node_compare(int a, int b)
470 return node_to_pxm(a) == node_to_pxm(b);
473 static int null_slit_node_compare(int a, int b)
477 #endif /* CONFIG_NUMA_EMU */
479 void __init srat_reserve_add_area(int nodeid)
481 if (found_add_area && nodes_add[nodeid].end) {
484 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
485 "for node %d at %Lx-%Lx\n",
486 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
487 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
489 total_mb *= sizeof(struct page);
491 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
492 "pre-allocated memory.\n", (unsigned long long)total_mb);
493 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
494 nodes_add[nodeid].end - nodes_add[nodeid].start,
499 int __node_distance(int a, int b)
504 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
506 index = acpi_slit->locality_count * node_to_pxm(a);
507 return acpi_slit->entry[index + node_to_pxm(b)];
510 EXPORT_SYMBOL(__node_distance);
512 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
513 int memory_add_physaddr_to_nid(u64 start)
518 if (nodes_add[i].start <= start && nodes_add[i].end > start)
523 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);