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 /* Callback for Proximity Domain -> LAPIC mapping */
108 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
115 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
119 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
121 pxm = pa->proximity_domain_lo;
122 node = setup_node(pxm);
124 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
130 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
132 apic_id = pa->apic_id;
133 apicid_to_node[apic_id] = node;
135 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
139 static int update_end_of_memory(unsigned long end) {return -1;}
140 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
141 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
142 static inline int save_add_info(void) {return 1;}
144 static inline int save_add_info(void) {return 0;}
147 * Update nodes_add and decide if to include add are in the zone.
148 * Both SPARSE and RESERVE need nodes_add information.
149 * This code supports one contiguous hot add area per node.
152 reserve_hotadd(int node, unsigned long start, unsigned long end)
154 unsigned long s_pfn = start >> PAGE_SHIFT;
155 unsigned long e_pfn = end >> PAGE_SHIFT;
156 int ret = 0, changed = 0;
157 struct bootnode *nd = &nodes_add[node];
159 /* I had some trouble with strange memory hotadd regions breaking
160 the boot. Be very strict here and reject anything unexpected.
161 If you want working memory hotadd write correct SRATs.
163 The node size check is a basic sanity check to guard against
165 if ((signed long)(end - start) < NODE_MIN_SIZE) {
166 printk(KERN_ERR "SRAT: Hotplug area too small\n");
170 /* This check might be a bit too strict, but I'm keeping it for now. */
171 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
173 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
178 if (!hotadd_enough_memory(&nodes_add[node])) {
179 printk(KERN_ERR "SRAT: Hotplug area too large\n");
185 if (nd->start == nd->end) {
190 if (nd->start == end) {
194 if (nd->end == start) {
199 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
202 ret = update_end_of_memory(nd->end);
205 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
209 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
211 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
213 struct bootnode *nd, oldnode;
214 unsigned long start, end;
220 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
224 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
227 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
229 start = ma->base_address;
230 end = start + ma->length;
231 pxm = ma->proximity_domain;
232 node = setup_node(pxm);
234 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
238 i = conflicting_memblks(start, end);
241 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
242 pxm, start, end, nodes[i].start, nodes[i].end);
245 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
246 pxm, start, end, node_to_pxm(i),
247 nodes[i].start, nodes[i].end);
253 if (!node_test_and_set(node, nodes_parsed)) {
257 if (start < nd->start)
263 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
265 e820_register_active_regions(node, start >> PAGE_SHIFT,
267 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
268 nd->end >> PAGE_SHIFT);
270 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
271 (reserve_hotadd(node, start, end) < 0)) {
272 /* Ignore hotadd region. Undo damage */
273 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
275 if ((nd->start | nd->end) == 0)
276 node_clear(node, nodes_parsed);
279 node_memblk_range[num_node_memblks].start = start;
280 node_memblk_range[num_node_memblks].end = end;
281 memblk_nodeid[num_node_memblks] = node;
285 /* Sanity check to catch more bad SRATs (they are amazingly common).
286 Make sure the PXMs cover all memory. */
287 static int __init nodes_cover_memory(const struct bootnode *nodes)
290 unsigned long pxmram, e820ram;
293 for_each_node_mask(i, nodes_parsed) {
294 unsigned long s = nodes[i].start >> PAGE_SHIFT;
295 unsigned long e = nodes[i].end >> PAGE_SHIFT;
297 pxmram -= absent_pages_in_range(s, e);
298 if ((long)pxmram < 0)
302 e820ram = end_pfn - absent_pages_in_range(0, end_pfn);
303 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
304 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
306 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
307 (pxmram << PAGE_SHIFT) >> 20,
308 (e820ram << PAGE_SHIFT) >> 20);
314 static void __init unparse_node(int node)
317 node_clear(node, nodes_parsed);
318 for (i = 0; i < MAX_LOCAL_APIC; i++) {
319 if (apicid_to_node[i] == node)
320 apicid_to_node[i] = NUMA_NO_NODE;
324 void __init acpi_numa_arch_fixup(void) {}
326 /* Use the information discovered above to actually set up the nodes. */
327 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
334 /* First clean up the node list */
335 for (i = 0; i < MAX_NUMNODES; i++) {
336 cutoff_node(i, start, end);
338 * don't confuse VM with a node that doesn't have the
342 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
348 if (!nodes_cover_memory(nodes)) {
353 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
355 if (memnode_shift < 0) {
357 "SRAT: No NUMA node hash function found. Contact maintainer\n");
362 node_possible_map = nodes_parsed;
364 /* Finally register nodes */
365 for_each_node_mask(i, node_possible_map)
366 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
367 /* Try again in case setup_node_bootmem missed one due
368 to missing bootmem */
369 for_each_node_mask(i, node_possible_map)
371 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
373 for (i = 0; i < NR_CPUS; i++) {
374 int node = early_cpu_to_node(i);
376 if (node == NUMA_NO_NODE)
378 if (!node_isset(node, node_possible_map))
385 #ifdef CONFIG_NUMA_EMU
386 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
387 [0 ... MAX_NUMNODES-1] = PXM_INVAL
389 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
390 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
392 static int __init find_node_by_addr(unsigned long addr)
394 int ret = NUMA_NO_NODE;
397 for_each_node_mask(i, nodes_parsed) {
399 * Find the real node that this emulated node appears on. For
400 * the sake of simplicity, we only use a real node's starting
401 * address to determine which emulated node it appears on.
403 if (addr >= nodes[i].start && addr < nodes[i].end) {
412 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
413 * mappings that respect the real ACPI topology but reflect our emulated
414 * environment. For each emulated node, we find which real node it appears on
415 * and create PXM to NID mappings for those fake nodes which mirror that
416 * locality. SLIT will now represent the correct distances between emulated
417 * nodes as a result of the real topology.
419 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
423 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
425 for (i = 0; i < num_nodes; i++) {
428 nid = find_node_by_addr(fake_nodes[i].start);
429 if (nid == NUMA_NO_NODE)
431 pxm = node_to_pxm(nid);
432 if (pxm == PXM_INVAL)
434 fake_node_to_pxm_map[i] = pxm;
436 * For each apicid_to_node mapping that exists for this real
437 * node, it must now point to the fake node ID.
439 for (j = 0; j < MAX_LOCAL_APIC; j++)
440 if (apicid_to_node[j] == nid)
441 fake_apicid_to_node[j] = i;
443 for (i = 0; i < num_nodes; i++)
444 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
445 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
447 nodes_clear(nodes_parsed);
448 for (i = 0; i < num_nodes; i++)
449 if (fake_nodes[i].start != fake_nodes[i].end)
450 node_set(i, nodes_parsed);
451 WARN_ON(!nodes_cover_memory(fake_nodes));
454 static int null_slit_node_compare(int a, int b)
456 return node_to_pxm(a) == node_to_pxm(b);
459 static int null_slit_node_compare(int a, int b)
463 #endif /* CONFIG_NUMA_EMU */
465 void __init srat_reserve_add_area(int nodeid)
467 if (found_add_area && nodes_add[nodeid].end) {
470 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
471 "for node %d at %Lx-%Lx\n",
472 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
473 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
475 total_mb *= sizeof(struct page);
477 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
478 "pre-allocated memory.\n", (unsigned long long)total_mb);
479 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
480 nodes_add[nodeid].end - nodes_add[nodeid].start,
485 int __node_distance(int a, int b)
490 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
492 index = acpi_slit->locality_count * node_to_pxm(a);
493 return acpi_slit->entry[index + node_to_pxm(b)];
496 EXPORT_SYMBOL(__node_distance);
498 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
499 int memory_add_physaddr_to_nid(u64 start)
504 if (nodes_add[i].start <= start && nodes_add[i].end > start)
509 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);