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 u16 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(u16
, x86_cpu_to_node_map
) = NUMA_NO_NODE
;
39 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map
);
40 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
41 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr
);
44 u16 apicid_to_node
[MAX_LOCAL_APIC
] __cpuinitdata
= {
45 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
48 cpumask_t node_to_cpumask_map
[MAX_NUMNODES
] __read_mostly
;
49 EXPORT_SYMBOL(node_to_cpumask_map
);
51 int numa_off __initdata
;
52 unsigned long __initdata nodemap_addr
;
53 unsigned long __initdata nodemap_size
;
56 * Given a shift value, try to populate memnodemap[]
59 * 0 if memnodmap[] too small (of shift too small)
60 * -1 if node overlap or lost ram (shift too big)
62 static int __init
populate_memnodemap(const struct bootnode
*nodes
,
63 int numnodes
, int shift
)
65 unsigned long addr
, end
;
68 memset(memnodemap
, 0xff, memnodemapsize
);
69 for (i
= 0; i
< numnodes
; i
++) {
70 addr
= nodes
[i
].start
;
74 if ((end
>> shift
) >= memnodemapsize
)
77 if (memnodemap
[addr
>> shift
] != 0xff)
79 memnodemap
[addr
>> shift
] = i
;
80 addr
+= (1UL << shift
);
87 static int __init
allocate_cachealigned_memnodemap(void)
89 unsigned long pad
, pad_addr
;
91 memnodemap
= memnode
.embedded_map
;
92 if (memnodemapsize
<= ARRAY_SIZE(memnode
.embedded_map
))
95 pad
= L1_CACHE_BYTES
- 1;
97 nodemap_size
= pad
+ memnodemapsize
;
98 nodemap_addr
= find_e820_area(pad_addr
, end_pfn
<<PAGE_SHIFT
,
100 if (nodemap_addr
== -1UL) {
102 "NUMA: Unable to allocate Memory to Node hash map\n");
103 nodemap_addr
= nodemap_size
= 0;
106 pad_addr
= (nodemap_addr
+ pad
) & ~pad
;
107 memnodemap
= phys_to_virt(pad_addr
);
108 reserve_early(nodemap_addr
, nodemap_addr
+ nodemap_size
);
110 printk(KERN_DEBUG
"NUMA: Allocated memnodemap from %lx - %lx\n",
111 nodemap_addr
, nodemap_addr
+ nodemap_size
);
116 * The LSB of all start and end addresses in the node map is the value of the
117 * maximum possible shift.
119 static int __init
extract_lsb_from_nodes(const struct bootnode
*nodes
,
122 int i
, nodes_used
= 0;
123 unsigned long start
, end
;
124 unsigned long bitfield
= 0, memtop
= 0;
126 for (i
= 0; i
< numnodes
; i
++) {
127 start
= nodes
[i
].start
;
139 i
= find_first_bit(&bitfield
, sizeof(unsigned long)*8);
140 memnodemapsize
= (memtop
>> i
)+1;
144 int __init
compute_hash_shift(struct bootnode
*nodes
, int numnodes
)
148 shift
= extract_lsb_from_nodes(nodes
, numnodes
);
149 if (allocate_cachealigned_memnodemap())
151 printk(KERN_DEBUG
"NUMA: Using %d for the hash shift.\n",
154 if (populate_memnodemap(nodes
, numnodes
, shift
) != 1) {
155 printk(KERN_INFO
"Your memory is not aligned you need to "
156 "rebuild your kernel with a bigger NODEMAPSIZE "
157 "shift=%d\n", shift
);
163 int early_pfn_to_nid(unsigned long pfn
)
165 return phys_to_nid(pfn
<< PAGE_SHIFT
);
168 static void * __init
early_node_mem(int nodeid
, unsigned long start
,
169 unsigned long end
, unsigned long size
)
171 unsigned long mem
= find_e820_area(start
, end
, size
);
176 ptr
= __alloc_bootmem_nopanic(size
,
177 SMP_CACHE_BYTES
, __pa(MAX_DMA_ADDRESS
));
179 printk(KERN_ERR
"Cannot find %lu bytes in node %d\n",
186 /* Initialize bootmem allocator for a node */
187 void __init
setup_node_bootmem(int nodeid
, unsigned long start
,
190 unsigned long start_pfn
, end_pfn
, bootmap_pages
, bootmap_size
;
191 unsigned long bootmap_start
, nodedata_phys
;
193 const int pgdat_size
= round_up(sizeof(pg_data_t
), PAGE_SIZE
);
195 start
= round_up(start
, ZONE_ALIGN
);
197 printk(KERN_INFO
"Bootmem setup node %d %016lx-%016lx\n", nodeid
,
200 start_pfn
= start
>> PAGE_SHIFT
;
201 end_pfn
= end
>> PAGE_SHIFT
;
203 node_data
[nodeid
] = early_node_mem(nodeid
, start
, end
, pgdat_size
);
204 if (node_data
[nodeid
] == NULL
)
206 nodedata_phys
= __pa(node_data
[nodeid
]);
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
);
216 bootmap
= early_node_mem(nodeid
, bootmap_start
, end
,
217 bootmap_pages
<<PAGE_SHIFT
);
218 if (bootmap
== NULL
) {
219 if (nodedata_phys
< start
|| nodedata_phys
>= end
)
220 free_bootmem((unsigned long)node_data
[nodeid
],
222 node_data
[nodeid
] = NULL
;
225 bootmap_start
= __pa(bootmap
);
226 Dprintk("bootmap start %lu pages %lu\n", bootmap_start
, bootmap_pages
);
228 bootmap_size
= init_bootmem_node(NODE_DATA(nodeid
),
229 bootmap_start
>> PAGE_SHIFT
,
232 free_bootmem_with_active_regions(nodeid
, end
);
234 reserve_bootmem_node(NODE_DATA(nodeid
), nodedata_phys
, pgdat_size
);
235 reserve_bootmem_node(NODE_DATA(nodeid
), bootmap_start
,
236 bootmap_pages
<<PAGE_SHIFT
);
237 #ifdef CONFIG_ACPI_NUMA
238 srat_reserve_add_area(nodeid
);
240 node_set_online(nodeid
);
244 * There are unfortunately some poorly designed mainboards around that
245 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
246 * mapping. To avoid this fill in the mapping for all possible CPUs,
247 * as the number of CPUs is not known yet. We round robin the existing
250 void __init
numa_init_array(void)
254 rr
= first_node(node_online_map
);
255 for (i
= 0; i
< NR_CPUS
; i
++) {
256 if (cpu_to_node(i
) != NUMA_NO_NODE
)
258 numa_set_node(i
, rr
);
259 rr
= next_node(rr
, node_online_map
);
260 if (rr
== MAX_NUMNODES
)
261 rr
= first_node(node_online_map
);
265 #ifdef CONFIG_NUMA_EMU
267 char *cmdline __initdata
;
270 * Setups up nid to range from addr to addr + size. If the end
271 * boundary is greater than max_addr, then max_addr is used instead.
272 * The return value is 0 if there is additional memory left for
273 * allocation past addr and -1 otherwise. addr is adjusted to be at
274 * the end of the node.
276 static int __init
setup_node_range(int nid
, struct bootnode
*nodes
, u64
*addr
,
277 u64 size
, u64 max_addr
)
281 nodes
[nid
].start
= *addr
;
283 if (*addr
>= max_addr
) {
287 nodes
[nid
].end
= *addr
;
288 node_set(nid
, node_possible_map
);
289 printk(KERN_INFO
"Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid
,
290 nodes
[nid
].start
, nodes
[nid
].end
,
291 (nodes
[nid
].end
- nodes
[nid
].start
) >> 20);
296 * Splits num_nodes nodes up equally starting at node_start. The return value
297 * is the number of nodes split up and addr is adjusted to be at the end of the
298 * last node allocated.
300 static int __init
split_nodes_equally(struct bootnode
*nodes
, u64
*addr
,
301 u64 max_addr
, int node_start
,
310 if (num_nodes
> MAX_NUMNODES
)
311 num_nodes
= MAX_NUMNODES
;
312 size
= (max_addr
- *addr
- e820_hole_size(*addr
, max_addr
)) /
315 * Calculate the number of big nodes that can be allocated as a result
316 * of consolidating the leftovers.
318 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * num_nodes
) /
321 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
322 size
&= FAKE_NODE_MIN_HASH_MASK
;
324 printk(KERN_ERR
"Not enough memory for each node. "
325 "NUMA emulation disabled.\n");
329 for (i
= node_start
; i
< num_nodes
+ node_start
; i
++) {
330 u64 end
= *addr
+ size
;
333 end
+= FAKE_NODE_MIN_SIZE
;
335 * The final node can have the remaining system RAM. Other
336 * nodes receive roughly the same amount of available pages.
338 if (i
== num_nodes
+ node_start
- 1)
341 while (end
- *addr
- e820_hole_size(*addr
, end
) <
343 end
+= FAKE_NODE_MIN_SIZE
;
344 if (end
> max_addr
) {
349 if (setup_node_range(i
, nodes
, addr
, end
- *addr
, max_addr
) < 0)
352 return i
- node_start
+ 1;
356 * Splits the remaining system RAM into chunks of size. The remaining memory is
357 * always assigned to a final node and can be asymmetric. Returns the number of
360 static int __init
split_nodes_by_size(struct bootnode
*nodes
, u64
*addr
,
361 u64 max_addr
, int node_start
, u64 size
)
364 size
= (size
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
365 while (!setup_node_range(i
++, nodes
, addr
, size
, max_addr
))
367 return i
- node_start
;
371 * Sets up the system RAM area from start_pfn to end_pfn according to the
372 * numa=fake command-line option.
374 static int __init
numa_emulation(unsigned long start_pfn
, unsigned long end_pfn
)
376 struct bootnode nodes
[MAX_NUMNODES
];
377 u64 size
, addr
= start_pfn
<< PAGE_SHIFT
;
378 u64 max_addr
= end_pfn
<< PAGE_SHIFT
;
379 int num_nodes
= 0, num
= 0, coeff_flag
, coeff
= -1, i
;
381 memset(&nodes
, 0, sizeof(nodes
));
383 * If the numa=fake command-line is just a single number N, split the
384 * system RAM into N fake nodes.
386 if (!strchr(cmdline
, '*') && !strchr(cmdline
, ',')) {
387 long n
= simple_strtol(cmdline
, NULL
, 0);
389 num_nodes
= split_nodes_equally(nodes
, &addr
, max_addr
, 0, n
);
395 /* Parse the command line. */
396 for (coeff_flag
= 0; ; cmdline
++) {
397 if (*cmdline
&& isdigit(*cmdline
)) {
398 num
= num
* 10 + *cmdline
- '0';
401 if (*cmdline
== '*') {
406 if (!*cmdline
|| *cmdline
== ',') {
410 * Round down to the nearest FAKE_NODE_MIN_SIZE.
411 * Command-line coefficients are in megabytes.
413 size
= ((u64
)num
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
415 for (i
= 0; i
< coeff
; i
++, num_nodes
++)
416 if (setup_node_range(num_nodes
, nodes
,
417 &addr
, size
, max_addr
) < 0)
429 /* Fill remainder of system RAM, if appropriate. */
430 if (addr
< max_addr
) {
431 if (coeff_flag
&& coeff
< 0) {
432 /* Split remaining nodes into num-sized chunks */
433 num_nodes
+= split_nodes_by_size(nodes
, &addr
, max_addr
,
437 switch (*(cmdline
- 1)) {
439 /* Split remaining nodes into coeff chunks */
442 num_nodes
+= split_nodes_equally(nodes
, &addr
, max_addr
,
446 /* Do not allocate remaining system RAM */
449 /* Give one final node */
450 setup_node_range(num_nodes
, nodes
, &addr
,
451 max_addr
- addr
, max_addr
);
456 memnode_shift
= compute_hash_shift(nodes
, num_nodes
);
457 if (memnode_shift
< 0) {
459 printk(KERN_ERR
"No NUMA hash function found. NUMA emulation "
465 * We need to vacate all active ranges that may have been registered by
466 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
467 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
469 remove_all_active_ranges();
470 #ifdef CONFIG_ACPI_NUMA
473 for_each_node_mask(i
, node_possible_map
) {
474 e820_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
475 nodes
[i
].end
>> PAGE_SHIFT
);
476 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
478 acpi_fake_nodes(nodes
, num_nodes
);
482 #endif /* CONFIG_NUMA_EMU */
484 void __init
numa_initmem_init(unsigned long start_pfn
, unsigned long end_pfn
)
488 nodes_clear(node_possible_map
);
490 #ifdef CONFIG_NUMA_EMU
491 if (cmdline
&& !numa_emulation(start_pfn
, end_pfn
))
493 nodes_clear(node_possible_map
);
496 #ifdef CONFIG_ACPI_NUMA
497 if (!numa_off
&& !acpi_scan_nodes(start_pfn
<< PAGE_SHIFT
,
498 end_pfn
<< PAGE_SHIFT
))
500 nodes_clear(node_possible_map
);
503 #ifdef CONFIG_K8_NUMA
504 if (!numa_off
&& !k8_scan_nodes(start_pfn
<<PAGE_SHIFT
,
505 end_pfn
<<PAGE_SHIFT
))
507 nodes_clear(node_possible_map
);
509 printk(KERN_INFO
"%s\n",
510 numa_off
? "NUMA turned off" : "No NUMA configuration found");
512 printk(KERN_INFO
"Faking a node at %016lx-%016lx\n",
513 start_pfn
<< PAGE_SHIFT
,
514 end_pfn
<< PAGE_SHIFT
);
515 /* setup dummy node covering all memory */
517 memnodemap
= memnode
.embedded_map
;
519 nodes_clear(node_online_map
);
521 node_set(0, node_possible_map
);
522 for (i
= 0; i
< NR_CPUS
; i
++)
524 /* cpumask_of_cpu() may not be available during early startup */
525 memset(&node_to_cpumask_map
[0], 0, sizeof(node_to_cpumask_map
[0]));
526 cpu_set(0, node_to_cpumask_map
[0]);
527 e820_register_active_regions(0, start_pfn
, end_pfn
);
528 setup_node_bootmem(0, start_pfn
<< PAGE_SHIFT
, end_pfn
<< PAGE_SHIFT
);
531 __cpuinit
void numa_add_cpu(int cpu
)
533 set_bit(cpu
, (unsigned long *)&node_to_cpumask_map
[cpu_to_node(cpu
)]);
536 void __cpuinit
numa_set_node(int cpu
, int node
)
538 u16
*cpu_to_node_map
= x86_cpu_to_node_map_early_ptr
;
540 cpu_pda(cpu
)->nodenumber
= node
;
543 cpu_to_node_map
[cpu
] = node
;
544 else if(per_cpu_offset(cpu
))
545 per_cpu(x86_cpu_to_node_map
, cpu
) = node
;
547 Dprintk(KERN_INFO
"Setting node for non-present cpu %d\n", cpu
);
550 unsigned long __init
numa_free_all_bootmem(void)
552 unsigned long pages
= 0;
555 for_each_online_node(i
)
556 pages
+= free_all_bootmem_node(NODE_DATA(i
));
561 void __init
paging_init(void)
563 unsigned long max_zone_pfns
[MAX_NR_ZONES
];
565 memset(max_zone_pfns
, 0, sizeof(max_zone_pfns
));
566 max_zone_pfns
[ZONE_DMA
] = MAX_DMA_PFN
;
567 max_zone_pfns
[ZONE_DMA32
] = MAX_DMA32_PFN
;
568 max_zone_pfns
[ZONE_NORMAL
] = end_pfn
;
570 sparse_memory_present_with_active_regions(MAX_NUMNODES
);
573 free_area_init_nodes(max_zone_pfns
);
576 static __init
int numa_setup(char *opt
)
580 if (!strncmp(opt
, "off", 3))
582 #ifdef CONFIG_NUMA_EMU
583 if (!strncmp(opt
, "fake=", 5))
586 #ifdef CONFIG_ACPI_NUMA
587 if (!strncmp(opt
, "noacpi", 6))
589 if (!strncmp(opt
, "hotadd=", 7))
590 hotadd_percent
= simple_strtoul(opt
+7, NULL
, 10);
594 early_param("numa", numa_setup
);
597 * Setup early cpu_to_node.
599 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
600 * and apicid_to_node[] tables have valid entries for a CPU.
601 * This means we skip cpu_to_node[] initialisation for NUMA
602 * emulation and faking node case (when running a kernel compiled
603 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
604 * is already initialized in a round robin manner at numa_init_array,
605 * prior to this call, and this initialization is good enough
606 * for the fake NUMA cases.
608 void __init
init_cpu_to_node(void)
612 for (i
= 0; i
< NR_CPUS
; i
++) {
613 u16 apicid
= x86_cpu_to_apicid_init
[i
];
615 if (apicid
== BAD_APICID
)
617 if (apicid_to_node
[apicid
] == NUMA_NO_NODE
)
619 numa_set_node(i
, apicid_to_node
[apicid
]);