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 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 -> x2APIC mapping */
121 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
128 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
132 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
134 pxm = pa->proximity_domain;
135 node = setup_node(pxm);
137 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
142 apic_id = pa->apic_id;
143 apicid_to_node[apic_id] = node;
145 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
149 /* Callback for Proximity Domain -> LAPIC mapping */
151 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
158 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
162 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
164 pxm = pa->proximity_domain_lo;
165 node = setup_node(pxm);
167 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
172 if (get_uv_system_type() >= UV_X2APIC)
173 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
175 apic_id = pa->apic_id;
176 apicid_to_node[apic_id] = node;
178 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
182 static int update_end_of_memory(unsigned long end) {return -1;}
183 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
184 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
185 static inline int save_add_info(void) {return 1;}
187 static inline int save_add_info(void) {return 0;}
190 * Update nodes_add and decide if to include add are in the zone.
191 * Both SPARSE and RESERVE need nodes_add information.
192 * This code supports one contiguous hot add area per node.
195 reserve_hotadd(int node, unsigned long start, unsigned long end)
197 unsigned long s_pfn = start >> PAGE_SHIFT;
198 unsigned long e_pfn = end >> PAGE_SHIFT;
199 int ret = 0, changed = 0;
200 struct bootnode *nd = &nodes_add[node];
202 /* I had some trouble with strange memory hotadd regions breaking
203 the boot. Be very strict here and reject anything unexpected.
204 If you want working memory hotadd write correct SRATs.
206 The node size check is a basic sanity check to guard against
208 if ((signed long)(end - start) < NODE_MIN_SIZE) {
209 printk(KERN_ERR "SRAT: Hotplug area too small\n");
213 /* This check might be a bit too strict, but I'm keeping it for now. */
214 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
216 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
221 if (!hotadd_enough_memory(&nodes_add[node])) {
222 printk(KERN_ERR "SRAT: Hotplug area too large\n");
228 if (nd->start == nd->end) {
233 if (nd->start == end) {
237 if (nd->end == start) {
242 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
245 ret = update_end_of_memory(nd->end);
248 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
252 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
254 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
256 struct bootnode *nd, oldnode;
257 unsigned long start, end;
263 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
267 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
270 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
272 start = ma->base_address;
273 end = start + ma->length;
274 pxm = ma->proximity_domain;
275 node = setup_node(pxm);
277 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
281 i = conflicting_memblks(start, end);
284 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
285 pxm, start, end, nodes[i].start, nodes[i].end);
288 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
289 pxm, start, end, node_to_pxm(i),
290 nodes[i].start, nodes[i].end);
296 if (!node_test_and_set(node, nodes_parsed)) {
300 if (start < nd->start)
306 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm,
308 e820_register_active_regions(node, start >> PAGE_SHIFT,
310 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
311 nd->end >> PAGE_SHIFT);
313 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
314 (reserve_hotadd(node, start, end) < 0)) {
315 /* Ignore hotadd region. Undo damage */
316 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
318 if ((nd->start | nd->end) == 0)
319 node_clear(node, nodes_parsed);
322 node_memblk_range[num_node_memblks].start = start;
323 node_memblk_range[num_node_memblks].end = end;
324 memblk_nodeid[num_node_memblks] = node;
328 /* Sanity check to catch more bad SRATs (they are amazingly common).
329 Make sure the PXMs cover all memory. */
330 static int __init nodes_cover_memory(const struct bootnode *nodes)
333 unsigned long pxmram, e820ram;
336 for_each_node_mask(i, nodes_parsed) {
337 unsigned long s = nodes[i].start >> PAGE_SHIFT;
338 unsigned long e = nodes[i].end >> PAGE_SHIFT;
340 pxmram -= absent_pages_in_range(s, e);
341 if ((long)pxmram < 0)
345 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
346 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
347 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
349 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
350 (pxmram << PAGE_SHIFT) >> 20,
351 (e820ram << PAGE_SHIFT) >> 20);
357 static void __init unparse_node(int node)
360 node_clear(node, nodes_parsed);
361 for (i = 0; i < MAX_LOCAL_APIC; i++) {
362 if (apicid_to_node[i] == node)
363 apicid_to_node[i] = NUMA_NO_NODE;
367 void __init acpi_numa_arch_fixup(void) {}
369 /* Use the information discovered above to actually set up the nodes. */
370 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
377 /* First clean up the node list */
378 for (i = 0; i < MAX_NUMNODES; i++) {
379 cutoff_node(i, start, end);
381 * don't confuse VM with a node that doesn't have the
385 (nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
391 if (!nodes_cover_memory(nodes)) {
396 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks,
398 if (memnode_shift < 0) {
400 "SRAT: No NUMA node hash function found. Contact maintainer\n");
405 node_possible_map = nodes_parsed;
407 /* Finally register nodes */
408 for_each_node_mask(i, node_possible_map)
409 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
410 /* Try again in case setup_node_bootmem missed one due
411 to missing bootmem */
412 for_each_node_mask(i, node_possible_map)
414 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
416 for (i = 0; i < nr_cpu_ids; i++) {
417 int node = early_cpu_to_node(i);
419 if (node == NUMA_NO_NODE)
421 if (!node_isset(node, node_possible_map))
428 #ifdef CONFIG_NUMA_EMU
429 static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = {
430 [0 ... MAX_NUMNODES-1] = PXM_INVAL
432 static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = {
433 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
435 static int __init find_node_by_addr(unsigned long addr)
437 int ret = NUMA_NO_NODE;
440 for_each_node_mask(i, nodes_parsed) {
442 * Find the real node that this emulated node appears on. For
443 * the sake of simplicity, we only use a real node's starting
444 * address to determine which emulated node it appears on.
446 if (addr >= nodes[i].start && addr < nodes[i].end) {
455 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
456 * mappings that respect the real ACPI topology but reflect our emulated
457 * environment. For each emulated node, we find which real node it appears on
458 * and create PXM to NID mappings for those fake nodes which mirror that
459 * locality. SLIT will now represent the correct distances between emulated
460 * nodes as a result of the real topology.
462 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
466 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
468 for (i = 0; i < num_nodes; i++) {
471 nid = find_node_by_addr(fake_nodes[i].start);
472 if (nid == NUMA_NO_NODE)
474 pxm = node_to_pxm(nid);
475 if (pxm == PXM_INVAL)
477 fake_node_to_pxm_map[i] = pxm;
479 * For each apicid_to_node mapping that exists for this real
480 * node, it must now point to the fake node ID.
482 for (j = 0; j < MAX_LOCAL_APIC; j++)
483 if (apicid_to_node[j] == nid)
484 fake_apicid_to_node[j] = i;
486 for (i = 0; i < num_nodes; i++)
487 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
488 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
490 nodes_clear(nodes_parsed);
491 for (i = 0; i < num_nodes; i++)
492 if (fake_nodes[i].start != fake_nodes[i].end)
493 node_set(i, nodes_parsed);
494 WARN_ON(!nodes_cover_memory(fake_nodes));
497 static int null_slit_node_compare(int a, int b)
499 return node_to_pxm(a) == node_to_pxm(b);
502 static int null_slit_node_compare(int a, int b)
506 #endif /* CONFIG_NUMA_EMU */
508 void __init srat_reserve_add_area(int nodeid)
510 if (found_add_area && nodes_add[nodeid].end) {
513 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
514 "for node %d at %Lx-%Lx\n",
515 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
516 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
518 total_mb *= sizeof(struct page);
520 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
521 "pre-allocated memory.\n", (unsigned long long)total_mb);
522 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
523 nodes_add[nodeid].end - nodes_add[nodeid].start,
528 int __node_distance(int a, int b)
533 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
535 index = acpi_slit->locality_count * node_to_pxm(a);
536 return acpi_slit->entry[index + node_to_pxm(b)];
539 EXPORT_SYMBOL(__node_distance);
541 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
542 int memory_add_physaddr_to_nid(u64 start)
547 if (nodes_add[i].start <= start && nodes_add[i].end > start)
552 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);