2 * Generic VM initialization for x86-64 NUMA setups.
3 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 #include <linux/kernel.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/mmzone.h>
11 #include <linux/ctype.h>
12 #include <linux/module.h>
13 #include <linux/nodemask.h>
14 #include <linux/sched.h>
17 #include <asm/proto.h>
27 struct pglist_data
*node_data
[MAX_NUMNODES
] __read_mostly
;
28 EXPORT_SYMBOL(node_data
);
30 bootmem_data_t plat_node_bdata
[MAX_NUMNODES
];
32 struct memnode memnode
;
34 int x86_cpu_to_node_map_init
[NR_CPUS
] = {
35 [0 ... NR_CPUS
-1] = NUMA_NO_NODE
37 void *x86_cpu_to_node_map_early_ptr
;
38 DEFINE_PER_CPU(int, x86_cpu_to_node_map
) = NUMA_NO_NODE
;
39 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map
);
40 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr
);
42 s16 apicid_to_node
[MAX_LOCAL_APIC
] __cpuinitdata
= {
43 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
46 cpumask_t node_to_cpumask_map
[MAX_NUMNODES
] __read_mostly
;
47 EXPORT_SYMBOL(node_to_cpumask_map
);
49 int numa_off __initdata
;
50 unsigned long __initdata nodemap_addr
;
51 unsigned long __initdata nodemap_size
;
54 * Given a shift value, try to populate memnodemap[]
57 * 0 if memnodmap[] too small (of shift too small)
58 * -1 if node overlap or lost ram (shift too big)
60 static int __init
populate_memnodemap(const struct bootnode
*nodes
,
61 int numnodes
, int shift
)
63 unsigned long addr
, end
;
66 memset(memnodemap
, 0xff, sizeof(s16
)*memnodemapsize
);
67 for (i
= 0; i
< numnodes
; i
++) {
68 addr
= nodes
[i
].start
;
72 if ((end
>> shift
) >= memnodemapsize
)
75 if (memnodemap
[addr
>> shift
] != NUMA_NO_NODE
)
77 memnodemap
[addr
>> shift
] = i
;
78 addr
+= (1UL << shift
);
85 static int __init
allocate_cachealigned_memnodemap(void)
89 memnodemap
= memnode
.embedded_map
;
90 if (memnodemapsize
<= ARRAY_SIZE(memnode
.embedded_map
))
94 nodemap_size
= round_up(sizeof(s16
) * memnodemapsize
, L1_CACHE_BYTES
);
95 nodemap_addr
= find_e820_area(addr
, end_pfn
<<PAGE_SHIFT
,
96 nodemap_size
, L1_CACHE_BYTES
);
97 if (nodemap_addr
== -1UL) {
99 "NUMA: Unable to allocate Memory to Node hash map\n");
100 nodemap_addr
= nodemap_size
= 0;
103 memnodemap
= phys_to_virt(nodemap_addr
);
104 reserve_early(nodemap_addr
, nodemap_addr
+ nodemap_size
, "MEMNODEMAP");
106 printk(KERN_DEBUG
"NUMA: Allocated memnodemap from %lx - %lx\n",
107 nodemap_addr
, nodemap_addr
+ nodemap_size
);
112 * The LSB of all start and end addresses in the node map is the value of the
113 * maximum possible shift.
115 static int __init
extract_lsb_from_nodes(const struct bootnode
*nodes
,
118 int i
, nodes_used
= 0;
119 unsigned long start
, end
;
120 unsigned long bitfield
= 0, memtop
= 0;
122 for (i
= 0; i
< numnodes
; i
++) {
123 start
= nodes
[i
].start
;
135 i
= find_first_bit(&bitfield
, sizeof(unsigned long)*8);
136 memnodemapsize
= (memtop
>> i
)+1;
140 int __init
compute_hash_shift(struct bootnode
*nodes
, int numnodes
)
144 shift
= extract_lsb_from_nodes(nodes
, numnodes
);
145 if (allocate_cachealigned_memnodemap())
147 printk(KERN_DEBUG
"NUMA: Using %d for the hash shift.\n",
150 if (populate_memnodemap(nodes
, numnodes
, shift
) != 1) {
151 printk(KERN_INFO
"Your memory is not aligned you need to "
152 "rebuild your kernel with a bigger NODEMAPSIZE "
153 "shift=%d\n", shift
);
159 int early_pfn_to_nid(unsigned long pfn
)
161 return phys_to_nid(pfn
<< PAGE_SHIFT
);
164 static void * __init
early_node_mem(int nodeid
, unsigned long start
,
165 unsigned long end
, unsigned long size
,
168 unsigned long mem
= find_e820_area(start
, end
, size
, align
);
174 ptr
= __alloc_bootmem_nopanic(size
, align
, __pa(MAX_DMA_ADDRESS
));
176 printk(KERN_ERR
"Cannot find %lu bytes in node %d\n",
183 /* Initialize bootmem allocator for a node */
184 void __init
setup_node_bootmem(int nodeid
, unsigned long start
,
187 unsigned long start_pfn
, end_pfn
, bootmap_pages
, bootmap_size
;
188 unsigned long bootmap_start
, nodedata_phys
;
190 const int pgdat_size
= round_up(sizeof(pg_data_t
), PAGE_SIZE
);
192 start
= round_up(start
, ZONE_ALIGN
);
194 printk(KERN_INFO
"Bootmem setup node %d %016lx-%016lx\n", nodeid
,
197 start_pfn
= start
>> PAGE_SHIFT
;
198 end_pfn
= end
>> PAGE_SHIFT
;
200 node_data
[nodeid
] = early_node_mem(nodeid
, start
, end
, pgdat_size
,
202 if (node_data
[nodeid
] == NULL
)
204 nodedata_phys
= __pa(node_data
[nodeid
]);
205 printk(KERN_INFO
" NODE_DATA [%016lx - %016lx]\n", nodedata_phys
,
206 nodedata_phys
+ pgdat_size
- 1);
208 memset(NODE_DATA(nodeid
), 0, sizeof(pg_data_t
));
209 NODE_DATA(nodeid
)->bdata
= &plat_node_bdata
[nodeid
];
210 NODE_DATA(nodeid
)->node_start_pfn
= start_pfn
;
211 NODE_DATA(nodeid
)->node_spanned_pages
= end_pfn
- start_pfn
;
213 /* Find a place for the bootmem map */
214 bootmap_pages
= bootmem_bootmap_pages(end_pfn
- start_pfn
);
215 bootmap_start
= round_up(nodedata_phys
+ pgdat_size
, PAGE_SIZE
);
217 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
218 * to use that to align to PAGE_SIZE
220 bootmap
= early_node_mem(nodeid
, bootmap_start
, end
,
221 bootmap_pages
<<PAGE_SHIFT
, PAGE_SIZE
);
222 if (bootmap
== NULL
) {
223 if (nodedata_phys
< start
|| nodedata_phys
>= end
)
224 free_bootmem(nodedata_phys
, pgdat_size
);
225 node_data
[nodeid
] = NULL
;
228 bootmap_start
= __pa(bootmap
);
230 bootmap_size
= init_bootmem_node(NODE_DATA(nodeid
),
231 bootmap_start
>> PAGE_SHIFT
,
234 printk(KERN_INFO
" bootmap [%016lx - %016lx] pages %lx\n",
235 bootmap_start
, bootmap_start
+ bootmap_size
- 1,
238 free_bootmem_with_active_regions(nodeid
, end
);
240 reserve_bootmem_node(NODE_DATA(nodeid
), nodedata_phys
, pgdat_size
,
242 reserve_bootmem_node(NODE_DATA(nodeid
), bootmap_start
,
243 bootmap_pages
<<PAGE_SHIFT
, BOOTMEM_DEFAULT
);
244 #ifdef CONFIG_ACPI_NUMA
245 srat_reserve_add_area(nodeid
);
247 node_set_online(nodeid
);
251 * There are unfortunately some poorly designed mainboards around that
252 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
253 * mapping. To avoid this fill in the mapping for all possible CPUs,
254 * as the number of CPUs is not known yet. We round robin the existing
257 void __init
numa_init_array(void)
261 rr
= first_node(node_online_map
);
262 for (i
= 0; i
< NR_CPUS
; i
++) {
263 if (early_cpu_to_node(i
) != NUMA_NO_NODE
)
265 numa_set_node(i
, rr
);
266 rr
= next_node(rr
, node_online_map
);
267 if (rr
== MAX_NUMNODES
)
268 rr
= first_node(node_online_map
);
272 #ifdef CONFIG_NUMA_EMU
274 char *cmdline __initdata
;
277 * Setups up nid to range from addr to addr + size. If the end
278 * boundary is greater than max_addr, then max_addr is used instead.
279 * The return value is 0 if there is additional memory left for
280 * allocation past addr and -1 otherwise. addr is adjusted to be at
281 * the end of the node.
283 static int __init
setup_node_range(int nid
, struct bootnode
*nodes
, u64
*addr
,
284 u64 size
, u64 max_addr
)
288 nodes
[nid
].start
= *addr
;
290 if (*addr
>= max_addr
) {
294 nodes
[nid
].end
= *addr
;
295 node_set(nid
, node_possible_map
);
296 printk(KERN_INFO
"Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid
,
297 nodes
[nid
].start
, nodes
[nid
].end
,
298 (nodes
[nid
].end
- nodes
[nid
].start
) >> 20);
303 * Splits num_nodes nodes up equally starting at node_start. The return value
304 * is the number of nodes split up and addr is adjusted to be at the end of the
305 * last node allocated.
307 static int __init
split_nodes_equally(struct bootnode
*nodes
, u64
*addr
,
308 u64 max_addr
, int node_start
,
317 if (num_nodes
> MAX_NUMNODES
)
318 num_nodes
= MAX_NUMNODES
;
319 size
= (max_addr
- *addr
- e820_hole_size(*addr
, max_addr
)) /
322 * Calculate the number of big nodes that can be allocated as a result
323 * of consolidating the leftovers.
325 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * num_nodes
) /
328 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
329 size
&= FAKE_NODE_MIN_HASH_MASK
;
331 printk(KERN_ERR
"Not enough memory for each node. "
332 "NUMA emulation disabled.\n");
336 for (i
= node_start
; i
< num_nodes
+ node_start
; i
++) {
337 u64 end
= *addr
+ size
;
340 end
+= FAKE_NODE_MIN_SIZE
;
342 * The final node can have the remaining system RAM. Other
343 * nodes receive roughly the same amount of available pages.
345 if (i
== num_nodes
+ node_start
- 1)
348 while (end
- *addr
- e820_hole_size(*addr
, end
) <
350 end
+= FAKE_NODE_MIN_SIZE
;
351 if (end
> max_addr
) {
356 if (setup_node_range(i
, nodes
, addr
, end
- *addr
, max_addr
) < 0)
359 return i
- node_start
+ 1;
363 * Splits the remaining system RAM into chunks of size. The remaining memory is
364 * always assigned to a final node and can be asymmetric. Returns the number of
367 static int __init
split_nodes_by_size(struct bootnode
*nodes
, u64
*addr
,
368 u64 max_addr
, int node_start
, u64 size
)
371 size
= (size
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
372 while (!setup_node_range(i
++, nodes
, addr
, size
, max_addr
))
374 return i
- node_start
;
378 * Sets up the system RAM area from start_pfn to end_pfn according to the
379 * numa=fake command-line option.
381 static int __init
numa_emulation(unsigned long start_pfn
, unsigned long end_pfn
)
383 struct bootnode nodes
[MAX_NUMNODES
];
384 u64 size
, addr
= start_pfn
<< PAGE_SHIFT
;
385 u64 max_addr
= end_pfn
<< PAGE_SHIFT
;
386 int num_nodes
= 0, num
= 0, coeff_flag
, coeff
= -1, i
;
388 memset(&nodes
, 0, sizeof(nodes
));
390 * If the numa=fake command-line is just a single number N, split the
391 * system RAM into N fake nodes.
393 if (!strchr(cmdline
, '*') && !strchr(cmdline
, ',')) {
394 long n
= simple_strtol(cmdline
, NULL
, 0);
396 num_nodes
= split_nodes_equally(nodes
, &addr
, max_addr
, 0, n
);
402 /* Parse the command line. */
403 for (coeff_flag
= 0; ; cmdline
++) {
404 if (*cmdline
&& isdigit(*cmdline
)) {
405 num
= num
* 10 + *cmdline
- '0';
408 if (*cmdline
== '*') {
413 if (!*cmdline
|| *cmdline
== ',') {
417 * Round down to the nearest FAKE_NODE_MIN_SIZE.
418 * Command-line coefficients are in megabytes.
420 size
= ((u64
)num
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
422 for (i
= 0; i
< coeff
; i
++, num_nodes
++)
423 if (setup_node_range(num_nodes
, nodes
,
424 &addr
, size
, max_addr
) < 0)
436 /* Fill remainder of system RAM, if appropriate. */
437 if (addr
< max_addr
) {
438 if (coeff_flag
&& coeff
< 0) {
439 /* Split remaining nodes into num-sized chunks */
440 num_nodes
+= split_nodes_by_size(nodes
, &addr
, max_addr
,
444 switch (*(cmdline
- 1)) {
446 /* Split remaining nodes into coeff chunks */
449 num_nodes
+= split_nodes_equally(nodes
, &addr
, max_addr
,
453 /* Do not allocate remaining system RAM */
456 /* Give one final node */
457 setup_node_range(num_nodes
, nodes
, &addr
,
458 max_addr
- addr
, max_addr
);
463 memnode_shift
= compute_hash_shift(nodes
, num_nodes
);
464 if (memnode_shift
< 0) {
466 printk(KERN_ERR
"No NUMA hash function found. NUMA emulation "
472 * We need to vacate all active ranges that may have been registered by
473 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
474 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
476 remove_all_active_ranges();
477 #ifdef CONFIG_ACPI_NUMA
480 for_each_node_mask(i
, node_possible_map
) {
481 e820_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
482 nodes
[i
].end
>> PAGE_SHIFT
);
483 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
485 acpi_fake_nodes(nodes
, num_nodes
);
489 #endif /* CONFIG_NUMA_EMU */
491 void __init
numa_initmem_init(unsigned long start_pfn
, unsigned long end_pfn
)
495 nodes_clear(node_possible_map
);
496 nodes_clear(node_online_map
);
498 #ifdef CONFIG_NUMA_EMU
499 if (cmdline
&& !numa_emulation(start_pfn
, end_pfn
))
501 nodes_clear(node_possible_map
);
502 nodes_clear(node_online_map
);
505 #ifdef CONFIG_ACPI_NUMA
506 if (!numa_off
&& !acpi_scan_nodes(start_pfn
<< PAGE_SHIFT
,
507 end_pfn
<< PAGE_SHIFT
))
509 nodes_clear(node_possible_map
);
510 nodes_clear(node_online_map
);
513 #ifdef CONFIG_K8_NUMA
514 if (!numa_off
&& !k8_scan_nodes(start_pfn
<<PAGE_SHIFT
,
515 end_pfn
<<PAGE_SHIFT
))
517 nodes_clear(node_possible_map
);
518 nodes_clear(node_online_map
);
520 printk(KERN_INFO
"%s\n",
521 numa_off
? "NUMA turned off" : "No NUMA configuration found");
523 printk(KERN_INFO
"Faking a node at %016lx-%016lx\n",
524 start_pfn
<< PAGE_SHIFT
,
525 end_pfn
<< PAGE_SHIFT
);
526 /* setup dummy node covering all memory */
528 memnodemap
= memnode
.embedded_map
;
531 node_set(0, node_possible_map
);
532 for (i
= 0; i
< NR_CPUS
; i
++)
534 /* cpumask_of_cpu() may not be available during early startup */
535 memset(&node_to_cpumask_map
[0], 0, sizeof(node_to_cpumask_map
[0]));
536 cpu_set(0, node_to_cpumask_map
[0]);
537 e820_register_active_regions(0, start_pfn
, end_pfn
);
538 setup_node_bootmem(0, start_pfn
<< PAGE_SHIFT
, end_pfn
<< PAGE_SHIFT
);
541 __cpuinit
void numa_add_cpu(int cpu
)
544 (unsigned long *)&node_to_cpumask_map
[early_cpu_to_node(cpu
)]);
547 void __cpuinit
numa_set_node(int cpu
, int node
)
549 int *cpu_to_node_map
= x86_cpu_to_node_map_early_ptr
;
552 cpu_to_node_map
[cpu
] = node
;
553 else if(per_cpu_offset(cpu
))
554 per_cpu(x86_cpu_to_node_map
, cpu
) = node
;
556 Dprintk(KERN_INFO
"Setting node for non-present cpu %d\n", cpu
);
559 unsigned long __init
numa_free_all_bootmem(void)
561 unsigned long pages
= 0;
564 for_each_online_node(i
)
565 pages
+= free_all_bootmem_node(NODE_DATA(i
));
570 void __init
paging_init(void)
572 unsigned long max_zone_pfns
[MAX_NR_ZONES
];
574 memset(max_zone_pfns
, 0, sizeof(max_zone_pfns
));
575 max_zone_pfns
[ZONE_DMA
] = MAX_DMA_PFN
;
576 max_zone_pfns
[ZONE_DMA32
] = MAX_DMA32_PFN
;
577 max_zone_pfns
[ZONE_NORMAL
] = end_pfn
;
579 sparse_memory_present_with_active_regions(MAX_NUMNODES
);
582 free_area_init_nodes(max_zone_pfns
);
585 static __init
int numa_setup(char *opt
)
589 if (!strncmp(opt
, "off", 3))
591 #ifdef CONFIG_NUMA_EMU
592 if (!strncmp(opt
, "fake=", 5))
595 #ifdef CONFIG_ACPI_NUMA
596 if (!strncmp(opt
, "noacpi", 6))
598 if (!strncmp(opt
, "hotadd=", 7))
599 hotadd_percent
= simple_strtoul(opt
+7, NULL
, 10);
603 early_param("numa", numa_setup
);
606 * Setup early cpu_to_node.
608 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
609 * and apicid_to_node[] tables have valid entries for a CPU.
610 * This means we skip cpu_to_node[] initialisation for NUMA
611 * emulation and faking node case (when running a kernel compiled
612 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
613 * is already initialized in a round robin manner at numa_init_array,
614 * prior to this call, and this initialization is good enough
615 * for the fake NUMA cases.
617 void __init
init_cpu_to_node(void)
621 for (i
= 0; i
< NR_CPUS
; i
++) {
623 u16 apicid
= x86_cpu_to_apicid_init
[i
];
625 if (apicid
== BAD_APICID
)
627 node
= apicid_to_node
[apicid
];
628 if (node
== NUMA_NO_NODE
)
630 if (!node_online(node
))
632 numa_set_node(i
, node
);