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 nodemask_t cpu_nodes_parsed __initdata
;
32 static struct bootnode nodes
[MAX_NUMNODES
] __initdata
;
33 static struct bootnode nodes_add
[MAX_NUMNODES
];
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 static __init
int setup_node(int pxm
)
41 return acpi_map_pxm_to_node(pxm
);
44 static __init
int conflicting_memblks(unsigned long start
, unsigned long end
)
47 for (i
= 0; i
< num_node_memblks
; i
++) {
48 struct bootnode
*nd
= &node_memblk_range
[i
];
49 if (nd
->start
== nd
->end
)
51 if (nd
->end
> start
&& nd
->start
< end
)
52 return memblk_nodeid
[i
];
53 if (nd
->end
== end
&& nd
->start
== start
)
54 return memblk_nodeid
[i
];
59 static __init
void cutoff_node(int i
, unsigned long start
, unsigned long end
)
61 struct bootnode
*nd
= &nodes
[i
];
63 if (nd
->start
< start
) {
65 if (nd
->end
< nd
->start
)
70 if (nd
->start
> nd
->end
)
75 static __init
void bad_srat(void)
78 printk(KERN_ERR
"SRAT: SRAT not used.\n");
80 for (i
= 0; i
< MAX_LOCAL_APIC
; i
++)
81 apicid_to_node
[i
] = NUMA_NO_NODE
;
82 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
83 nodes
[i
].start
= nodes
[i
].end
= 0;
84 nodes_add
[i
].start
= nodes_add
[i
].end
= 0;
86 remove_all_active_ranges();
89 static __init
inline int srat_disabled(void)
91 return numa_off
|| acpi_numa
< 0;
94 /* Callback for SLIT parsing */
95 void __init
acpi_numa_slit_init(struct acpi_table_slit
*slit
)
100 length
= slit
->header
.length
;
101 phys
= find_e820_area(0, max_pfn_mapped
<<PAGE_SHIFT
, length
,
105 panic(" Can not save slit!\n");
107 acpi_slit
= __va(phys
);
108 memcpy(acpi_slit
, slit
, length
);
109 reserve_early(phys
, phys
+ length
, "ACPI SLIT");
112 /* Callback for Proximity Domain -> x2APIC mapping */
114 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity
*pa
)
121 if (pa
->header
.length
< sizeof(struct acpi_srat_x2apic_cpu_affinity
)) {
125 if ((pa
->flags
& ACPI_SRAT_CPU_ENABLED
) == 0)
127 pxm
= pa
->proximity_domain
;
128 node
= setup_node(pxm
);
130 printk(KERN_ERR
"SRAT: Too many proximity domains %x\n", pxm
);
135 apic_id
= pa
->apic_id
;
136 apicid_to_node
[apic_id
] = node
;
137 node_set(node
, cpu_nodes_parsed
);
139 printk(KERN_INFO
"SRAT: PXM %u -> APIC %u -> Node %u\n",
143 /* Callback for Proximity Domain -> LAPIC mapping */
145 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity
*pa
)
152 if (pa
->header
.length
!= sizeof(struct acpi_srat_cpu_affinity
)) {
156 if ((pa
->flags
& ACPI_SRAT_CPU_ENABLED
) == 0)
158 pxm
= pa
->proximity_domain_lo
;
159 node
= setup_node(pxm
);
161 printk(KERN_ERR
"SRAT: Too many proximity domains %x\n", pxm
);
166 if (get_uv_system_type() >= UV_X2APIC
)
167 apic_id
= (pa
->apic_id
<< 8) | pa
->local_sapic_eid
;
169 apic_id
= pa
->apic_id
;
170 apicid_to_node
[apic_id
] = node
;
171 node_set(node
, cpu_nodes_parsed
);
173 printk(KERN_INFO
"SRAT: PXM %u -> APIC %u -> Node %u\n",
177 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
178 static inline int save_add_info(void) {return 1;}
180 static inline int save_add_info(void) {return 0;}
184 * This code supports one contiguous hot add area per node
187 update_nodes_add(int node
, unsigned long start
, unsigned long end
)
189 unsigned long s_pfn
= start
>> PAGE_SHIFT
;
190 unsigned long e_pfn
= end
>> PAGE_SHIFT
;
192 struct bootnode
*nd
= &nodes_add
[node
];
194 /* I had some trouble with strange memory hotadd regions breaking
195 the boot. Be very strict here and reject anything unexpected.
196 If you want working memory hotadd write correct SRATs.
198 The node size check is a basic sanity check to guard against
200 if ((signed long)(end
- start
) < NODE_MIN_SIZE
) {
201 printk(KERN_ERR
"SRAT: Hotplug area too small\n");
205 /* This check might be a bit too strict, but I'm keeping it for now. */
206 if (absent_pages_in_range(s_pfn
, e_pfn
) != e_pfn
- s_pfn
) {
208 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
215 if (nd
->start
== nd
->end
) {
220 if (nd
->start
== end
) {
224 if (nd
->end
== start
) {
229 printk(KERN_ERR
"SRAT: Hotplug zone not continuous. Partly ignored\n");
233 printk(KERN_INFO
"SRAT: hot plug zone found %Lx - %Lx\n",
237 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
239 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity
*ma
)
241 struct bootnode
*nd
, oldnode
;
242 unsigned long start
, end
;
248 if (ma
->header
.length
!= sizeof(struct acpi_srat_mem_affinity
)) {
252 if ((ma
->flags
& ACPI_SRAT_MEM_ENABLED
) == 0)
255 if ((ma
->flags
& ACPI_SRAT_MEM_HOT_PLUGGABLE
) && !save_add_info())
257 start
= ma
->base_address
;
258 end
= start
+ ma
->length
;
259 pxm
= ma
->proximity_domain
;
260 node
= setup_node(pxm
);
262 printk(KERN_ERR
"SRAT: Too many proximity domains.\n");
266 i
= conflicting_memblks(start
, end
);
269 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
270 pxm
, start
, end
, nodes
[i
].start
, nodes
[i
].end
);
273 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
274 pxm
, start
, end
, node_to_pxm(i
),
275 nodes
[i
].start
, nodes
[i
].end
);
281 if (!node_test_and_set(node
, nodes_parsed
)) {
285 if (start
< nd
->start
)
291 printk(KERN_INFO
"SRAT: Node %u PXM %u %lx-%lx\n", node
, pxm
,
294 if (ma
->flags
& ACPI_SRAT_MEM_HOT_PLUGGABLE
) {
295 update_nodes_add(node
, start
, end
);
296 /* restore nodes[node] */
298 if ((nd
->start
| nd
->end
) == 0)
299 node_clear(node
, nodes_parsed
);
302 node_memblk_range
[num_node_memblks
].start
= start
;
303 node_memblk_range
[num_node_memblks
].end
= end
;
304 memblk_nodeid
[num_node_memblks
] = node
;
308 /* Sanity check to catch more bad SRATs (they are amazingly common).
309 Make sure the PXMs cover all memory. */
310 static int __init
nodes_cover_memory(const struct bootnode
*nodes
)
313 unsigned long pxmram
, e820ram
;
316 for_each_node_mask(i
, nodes_parsed
) {
317 unsigned long s
= nodes
[i
].start
>> PAGE_SHIFT
;
318 unsigned long e
= nodes
[i
].end
>> PAGE_SHIFT
;
320 pxmram
-= absent_pages_in_range(s
, e
);
321 if ((long)pxmram
< 0)
325 e820ram
= max_pfn
- (e820_hole_size(0, max_pfn
<<PAGE_SHIFT
)>>PAGE_SHIFT
);
326 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
327 if ((long)(e820ram
- pxmram
) >= (1<<(20 - PAGE_SHIFT
))) {
329 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
330 (pxmram
<< PAGE_SHIFT
) >> 20,
331 (e820ram
<< PAGE_SHIFT
) >> 20);
337 void __init
acpi_numa_arch_fixup(void) {}
339 int __init
acpi_get_nodes(struct bootnode
*physnodes
)
344 for_each_node_mask(i
, nodes_parsed
) {
345 physnodes
[ret
].start
= nodes
[i
].start
;
346 physnodes
[ret
].end
= nodes
[i
].end
;
352 /* Use the information discovered above to actually set up the nodes. */
353 int __init
acpi_scan_nodes(unsigned long start
, unsigned long end
)
360 /* First clean up the node list */
361 for (i
= 0; i
< MAX_NUMNODES
; i
++)
362 cutoff_node(i
, start
, end
);
364 memnode_shift
= compute_hash_shift(node_memblk_range
, num_node_memblks
,
366 if (memnode_shift
< 0) {
368 "SRAT: No NUMA node hash function found. Contact maintainer\n");
373 for_each_node_mask(i
, nodes_parsed
)
374 e820_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
375 nodes
[i
].end
>> PAGE_SHIFT
);
376 if (!nodes_cover_memory(nodes
)) {
381 /* Account for nodes with cpus and no memory */
382 nodes_or(node_possible_map
, nodes_parsed
, cpu_nodes_parsed
);
384 /* Finally register nodes */
385 for_each_node_mask(i
, node_possible_map
)
386 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
387 /* Try again in case setup_node_bootmem missed one due
388 to missing bootmem */
389 for_each_node_mask(i
, node_possible_map
)
391 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
393 for (i
= 0; i
< nr_cpu_ids
; i
++) {
394 int node
= early_cpu_to_node(i
);
396 if (node
== NUMA_NO_NODE
)
398 if (!node_online(node
))
405 #ifdef CONFIG_NUMA_EMU
406 static int fake_node_to_pxm_map
[MAX_NUMNODES
] __initdata
= {
407 [0 ... MAX_NUMNODES
-1] = PXM_INVAL
409 static s16 fake_apicid_to_node
[MAX_LOCAL_APIC
] __initdata
= {
410 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
412 static int __init
find_node_by_addr(unsigned long addr
)
414 int ret
= NUMA_NO_NODE
;
417 for_each_node_mask(i
, nodes_parsed
) {
419 * Find the real node that this emulated node appears on. For
420 * the sake of simplicity, we only use a real node's starting
421 * address to determine which emulated node it appears on.
423 if (addr
>= nodes
[i
].start
&& addr
< nodes
[i
].end
) {
432 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
433 * mappings that respect the real ACPI topology but reflect our emulated
434 * environment. For each emulated node, we find which real node it appears on
435 * and create PXM to NID mappings for those fake nodes which mirror that
436 * locality. SLIT will now represent the correct distances between emulated
437 * nodes as a result of the real topology.
439 void __init
acpi_fake_nodes(const struct bootnode
*fake_nodes
, int num_nodes
)
443 printk(KERN_INFO
"Faking PXM affinity for fake nodes on real "
445 for (i
= 0; i
< num_nodes
; i
++) {
448 nid
= find_node_by_addr(fake_nodes
[i
].start
);
449 if (nid
== NUMA_NO_NODE
)
451 pxm
= node_to_pxm(nid
);
452 if (pxm
== PXM_INVAL
)
454 fake_node_to_pxm_map
[i
] = pxm
;
456 * For each apicid_to_node mapping that exists for this real
457 * node, it must now point to the fake node ID.
459 for (j
= 0; j
< MAX_LOCAL_APIC
; j
++)
460 if (apicid_to_node
[j
] == nid
)
461 fake_apicid_to_node
[j
] = i
;
463 for (i
= 0; i
< num_nodes
; i
++)
464 __acpi_map_pxm_to_node(fake_node_to_pxm_map
[i
], i
);
465 memcpy(apicid_to_node
, fake_apicid_to_node
, sizeof(apicid_to_node
));
467 nodes_clear(nodes_parsed
);
468 for (i
= 0; i
< num_nodes
; i
++)
469 if (fake_nodes
[i
].start
!= fake_nodes
[i
].end
)
470 node_set(i
, nodes_parsed
);
471 WARN_ON(!nodes_cover_memory(fake_nodes
));
474 static int null_slit_node_compare(int a
, int b
)
476 return node_to_pxm(a
) == node_to_pxm(b
);
479 static int null_slit_node_compare(int a
, int b
)
483 #endif /* CONFIG_NUMA_EMU */
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
);