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>
19 #include <linux/memblock.h>
21 #include <asm/proto.h>
25 #include <asm/uv/uv.h>
27 int acpi_numa __initdata
;
29 static struct acpi_table_slit
*acpi_slit
;
31 static struct bootnode nodes_add
[MAX_NUMNODES
];
33 static __init
int setup_node(int pxm
)
35 return acpi_map_pxm_to_node(pxm
);
38 static __init
void bad_srat(void)
41 printk(KERN_ERR
"SRAT: SRAT not used.\n");
43 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
44 numa_nodes
[i
].start
= numa_nodes
[i
].end
= 0;
45 nodes_add
[i
].start
= nodes_add
[i
].end
= 0;
47 remove_all_active_ranges();
50 static __init
inline int srat_disabled(void)
55 /* Callback for SLIT parsing */
56 void __init
acpi_numa_slit_init(struct acpi_table_slit
*slit
)
61 length
= slit
->header
.length
;
62 phys
= memblock_find_in_range(0, max_pfn_mapped
<<PAGE_SHIFT
, length
,
65 if (phys
== MEMBLOCK_ERROR
)
66 panic(" Can not save slit!\n");
68 acpi_slit
= __va(phys
);
69 memcpy(acpi_slit
, slit
, length
);
70 memblock_x86_reserve_range(phys
, phys
+ length
, "ACPI SLIT");
73 /* Callback for Proximity Domain -> x2APIC mapping */
75 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity
*pa
)
82 if (pa
->header
.length
< sizeof(struct acpi_srat_x2apic_cpu_affinity
)) {
86 if ((pa
->flags
& ACPI_SRAT_CPU_ENABLED
) == 0)
88 pxm
= pa
->proximity_domain
;
89 node
= setup_node(pxm
);
91 printk(KERN_ERR
"SRAT: Too many proximity domains %x\n", pxm
);
96 apic_id
= pa
->apic_id
;
97 if (apic_id
>= MAX_LOCAL_APIC
) {
98 printk(KERN_INFO
"SRAT: PXM %u -> APIC 0x%04x -> Node %u skipped apicid that is too big\n", pxm
, apic_id
, node
);
101 set_apicid_to_node(apic_id
, node
);
102 node_set(node
, cpu_nodes_parsed
);
104 printk(KERN_INFO
"SRAT: PXM %u -> APIC 0x%04x -> Node %u\n",
108 /* Callback for Proximity Domain -> LAPIC mapping */
110 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity
*pa
)
117 if (pa
->header
.length
!= sizeof(struct acpi_srat_cpu_affinity
)) {
121 if ((pa
->flags
& ACPI_SRAT_CPU_ENABLED
) == 0)
123 pxm
= pa
->proximity_domain_lo
;
124 node
= setup_node(pxm
);
126 printk(KERN_ERR
"SRAT: Too many proximity domains %x\n", pxm
);
131 if (get_uv_system_type() >= UV_X2APIC
)
132 apic_id
= (pa
->apic_id
<< 8) | pa
->local_sapic_eid
;
134 apic_id
= pa
->apic_id
;
136 if (apic_id
>= MAX_LOCAL_APIC
) {
137 printk(KERN_INFO
"SRAT: PXM %u -> APIC 0x%02x -> Node %u skipped apicid that is too big\n", pxm
, apic_id
, node
);
141 set_apicid_to_node(apic_id
, node
);
142 node_set(node
, cpu_nodes_parsed
);
144 printk(KERN_INFO
"SRAT: PXM %u -> APIC 0x%02x -> Node %u\n",
148 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
149 static inline int save_add_info(void) {return 1;}
151 static inline int save_add_info(void) {return 0;}
155 * This code supports one contiguous hot add area per node
158 update_nodes_add(int node
, unsigned long start
, unsigned long end
)
160 unsigned long s_pfn
= start
>> PAGE_SHIFT
;
161 unsigned long e_pfn
= end
>> PAGE_SHIFT
;
163 struct bootnode
*nd
= &nodes_add
[node
];
165 /* I had some trouble with strange memory hotadd regions breaking
166 the boot. Be very strict here and reject anything unexpected.
167 If you want working memory hotadd write correct SRATs.
169 The node size check is a basic sanity check to guard against
171 if ((signed long)(end
- start
) < NODE_MIN_SIZE
) {
172 printk(KERN_ERR
"SRAT: Hotplug area too small\n");
176 /* This check might be a bit too strict, but I'm keeping it for now. */
177 if (absent_pages_in_range(s_pfn
, e_pfn
) != e_pfn
- s_pfn
) {
179 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
186 if (nd
->start
== nd
->end
) {
191 if (nd
->start
== end
) {
195 if (nd
->end
== start
) {
200 printk(KERN_ERR
"SRAT: Hotplug zone not continuous. Partly ignored\n");
204 node_set(node
, cpu_nodes_parsed
);
205 printk(KERN_INFO
"SRAT: hot plug zone found %Lx - %Lx\n",
210 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
212 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity
*ma
)
215 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");
239 if (numa_add_memblk(node
, start
, end
) < 0) {
244 printk(KERN_INFO
"SRAT: Node %u PXM %u %lx-%lx\n", node
, pxm
,
247 if (!(ma
->flags
& ACPI_SRAT_MEM_HOT_PLUGGABLE
)) {
248 nd
= &numa_nodes
[node
];
249 if (!node_test_and_set(node
, mem_nodes_parsed
)) {
253 if (start
< nd
->start
)
259 update_nodes_add(node
, start
, end
);
262 /* Sanity check to catch more bad SRATs (they are amazingly common).
263 Make sure the PXMs cover all memory. */
264 static int __init
nodes_cover_memory(const struct bootnode
*nodes
)
267 unsigned long pxmram
, e820ram
;
270 for_each_node_mask(i
, mem_nodes_parsed
) {
271 unsigned long s
= nodes
[i
].start
>> PAGE_SHIFT
;
272 unsigned long e
= nodes
[i
].end
>> PAGE_SHIFT
;
274 pxmram
-= __absent_pages_in_range(i
, s
, e
);
275 if ((long)pxmram
< 0)
279 e820ram
= max_pfn
- (memblock_x86_hole_size(0, max_pfn
<<PAGE_SHIFT
)>>PAGE_SHIFT
);
280 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
281 if ((long)(e820ram
- pxmram
) >= (1<<(20 - PAGE_SHIFT
))) {
283 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
284 (pxmram
<< PAGE_SHIFT
) >> 20,
285 (e820ram
<< PAGE_SHIFT
) >> 20);
291 void __init
acpi_numa_arch_fixup(void) {}
293 int __init
x86_acpi_numa_init(void)
297 ret
= acpi_numa_init();
300 return srat_disabled() ? -EINVAL
: 0;
303 /* Use the information discovered above to actually set up the nodes. */
304 int __init
acpi_scan_nodes(void)
311 if (numa_register_memblks() < 0) {
316 /* for out of order entries in SRAT */
318 if (!nodes_cover_memory(numa_nodes
)) {
323 init_memory_mapping_high();
325 /* Finally register nodes */
326 for_each_node_mask(i
, node_possible_map
)
327 setup_node_bootmem(i
, numa_nodes
[i
].start
, numa_nodes
[i
].end
);
328 /* Try again in case setup_node_bootmem missed one due
329 to missing bootmem */
330 for_each_node_mask(i
, node_possible_map
)
332 setup_node_bootmem(i
, numa_nodes
[i
].start
,
335 for (i
= 0; i
< nr_cpu_ids
; i
++) {
336 int node
= early_cpu_to_node(i
);
338 if (node
== NUMA_NO_NODE
)
340 if (!node_online(node
))
347 #ifdef CONFIG_NUMA_EMU
348 static int fake_node_to_pxm_map
[MAX_NUMNODES
] __initdata
= {
349 [0 ... MAX_NUMNODES
-1] = PXM_INVAL
351 static s16 fake_apicid_to_node
[MAX_LOCAL_APIC
] __initdata
= {
352 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
354 static int __init
find_node_by_addr(unsigned long addr
)
356 int ret
= NUMA_NO_NODE
;
359 for_each_node_mask(i
, mem_nodes_parsed
) {
361 * Find the real node that this emulated node appears on. For
362 * the sake of simplicity, we only use a real node's starting
363 * address to determine which emulated node it appears on.
365 if (addr
>= numa_nodes
[i
].start
&& addr
< numa_nodes
[i
].end
) {
374 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
375 * mappings that respect the real ACPI topology but reflect our emulated
376 * environment. For each emulated node, we find which real node it appears on
377 * and create PXM to NID mappings for those fake nodes which mirror that
378 * locality. SLIT will now represent the correct distances between emulated
379 * nodes as a result of the real topology.
381 void __init
acpi_fake_nodes(const struct bootnode
*fake_nodes
, int num_nodes
)
385 for (i
= 0; i
< num_nodes
; i
++) {
388 nid
= find_node_by_addr(fake_nodes
[i
].start
);
389 if (nid
== NUMA_NO_NODE
)
391 pxm
= node_to_pxm(nid
);
392 if (pxm
== PXM_INVAL
)
394 fake_node_to_pxm_map
[i
] = pxm
;
396 * For each apicid_to_node mapping that exists for this real
397 * node, it must now point to the fake node ID.
399 for (j
= 0; j
< MAX_LOCAL_APIC
; j
++)
400 if (__apicid_to_node
[j
] == nid
&&
401 fake_apicid_to_node
[j
] == NUMA_NO_NODE
)
402 fake_apicid_to_node
[j
] = i
;
406 * If there are apicid-to-node mappings for physical nodes that do not
407 * have a corresponding emulated node, it should default to a guaranteed
410 for (i
= 0; i
< MAX_LOCAL_APIC
; i
++)
411 if (__apicid_to_node
[i
] != NUMA_NO_NODE
&&
412 fake_apicid_to_node
[i
] == NUMA_NO_NODE
)
413 fake_apicid_to_node
[i
] = 0;
415 for (i
= 0; i
< num_nodes
; i
++)
416 __acpi_map_pxm_to_node(fake_node_to_pxm_map
[i
], i
);
417 memcpy(__apicid_to_node
, fake_apicid_to_node
, sizeof(__apicid_to_node
));
419 nodes_clear(mem_nodes_parsed
);
420 for (i
= 0; i
< num_nodes
; i
++)
421 if (fake_nodes
[i
].start
!= fake_nodes
[i
].end
)
422 node_set(i
, mem_nodes_parsed
);
425 static int null_slit_node_compare(int a
, int b
)
427 return node_to_pxm(a
) == node_to_pxm(b
);
430 static int null_slit_node_compare(int a
, int b
)
434 #endif /* CONFIG_NUMA_EMU */
436 int __node_distance(int a
, int b
)
441 return null_slit_node_compare(a
, b
) ? LOCAL_DISTANCE
:
443 index
= acpi_slit
->locality_count
* node_to_pxm(a
);
444 return acpi_slit
->entry
[index
+ node_to_pxm(b
)];
447 EXPORT_SYMBOL(__node_distance
);
449 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY)
450 int memory_add_physaddr_to_nid(u64 start
)
455 if (nodes_add
[i
].start
<= start
&& nodes_add
[i
].end
> start
)
460 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid
);