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 cpu_to_node_map
[NR_CPUS
] __read_mostly
= {
35 [0 ... NR_CPUS
-1] = NUMA_NO_NODE
37 EXPORT_SYMBOL(cpu_to_node_map
);
39 u16 apicid_to_node
[MAX_LOCAL_APIC
] __cpuinitdata
= {
40 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
43 cpumask_t node_to_cpumask_map
[MAX_NUMNODES
] __read_mostly
;
44 EXPORT_SYMBOL(node_to_cpumask_map
);
46 int numa_off __initdata
;
47 unsigned long __initdata nodemap_addr
;
48 unsigned long __initdata nodemap_size
;
51 * Given a shift value, try to populate memnodemap[]
54 * 0 if memnodmap[] too small (of shift too small)
55 * -1 if node overlap or lost ram (shift too big)
57 static int __init
populate_memnodemap(const struct bootnode
*nodes
,
58 int numnodes
, int shift
)
60 unsigned long addr
, end
;
63 memset(memnodemap
, 0xff, memnodemapsize
);
64 for (i
= 0; i
< numnodes
; i
++) {
65 addr
= nodes
[i
].start
;
69 if ((end
>> shift
) >= memnodemapsize
)
72 if (memnodemap
[addr
>> shift
] != 0xff)
74 memnodemap
[addr
>> shift
] = i
;
75 addr
+= (1UL << shift
);
82 static int __init
allocate_cachealigned_memnodemap(void)
84 unsigned long pad
, pad_addr
;
86 memnodemap
= memnode
.embedded_map
;
87 if (memnodemapsize
<= 48)
90 pad
= L1_CACHE_BYTES
- 1;
92 nodemap_size
= pad
+ memnodemapsize
;
93 nodemap_addr
= find_e820_area(pad_addr
, end_pfn
<<PAGE_SHIFT
,
95 if (nodemap_addr
== -1UL) {
97 "NUMA: Unable to allocate Memory to Node hash map\n");
98 nodemap_addr
= nodemap_size
= 0;
101 pad_addr
= (nodemap_addr
+ pad
) & ~pad
;
102 memnodemap
= phys_to_virt(pad_addr
);
104 printk(KERN_DEBUG
"NUMA: Allocated memnodemap from %lx - %lx\n",
105 nodemap_addr
, nodemap_addr
+ nodemap_size
);
110 * The LSB of all start and end addresses in the node map is the value of the
111 * maximum possible shift.
113 static int __init
extract_lsb_from_nodes(const struct bootnode
*nodes
,
116 int i
, nodes_used
= 0;
117 unsigned long start
, end
;
118 unsigned long bitfield
= 0, memtop
= 0;
120 for (i
= 0; i
< numnodes
; i
++) {
121 start
= nodes
[i
].start
;
133 i
= find_first_bit(&bitfield
, sizeof(unsigned long)*8);
134 memnodemapsize
= (memtop
>> i
)+1;
138 int __init
compute_hash_shift(struct bootnode
*nodes
, int numnodes
)
142 shift
= extract_lsb_from_nodes(nodes
, numnodes
);
143 if (allocate_cachealigned_memnodemap())
145 printk(KERN_DEBUG
"NUMA: Using %d for the hash shift.\n",
148 if (populate_memnodemap(nodes
, numnodes
, shift
) != 1) {
149 printk(KERN_INFO
"Your memory is not aligned you need to "
150 "rebuild your kernel with a bigger NODEMAPSIZE "
151 "shift=%d\n", shift
);
157 int early_pfn_to_nid(unsigned long pfn
)
159 return phys_to_nid(pfn
<< PAGE_SHIFT
);
162 static void * __init
early_node_mem(int nodeid
, unsigned long start
,
163 unsigned long end
, unsigned long size
)
165 unsigned long mem
= find_e820_area(start
, end
, size
);
170 ptr
= __alloc_bootmem_nopanic(size
,
171 SMP_CACHE_BYTES
, __pa(MAX_DMA_ADDRESS
));
173 printk(KERN_ERR
"Cannot find %lu bytes in node %d\n",
180 /* Initialize bootmem allocator for a node */
181 void __init
setup_node_bootmem(int nodeid
, unsigned long start
,
184 unsigned long start_pfn
, end_pfn
, bootmap_pages
, bootmap_size
;
185 unsigned long bootmap_start
, nodedata_phys
;
187 const int pgdat_size
= round_up(sizeof(pg_data_t
), PAGE_SIZE
);
189 start
= round_up(start
, ZONE_ALIGN
);
191 printk(KERN_INFO
"Bootmem setup node %d %016lx-%016lx\n", nodeid
,
194 start_pfn
= start
>> PAGE_SHIFT
;
195 end_pfn
= end
>> PAGE_SHIFT
;
197 node_data
[nodeid
] = early_node_mem(nodeid
, start
, end
, pgdat_size
);
198 if (node_data
[nodeid
] == NULL
)
200 nodedata_phys
= __pa(node_data
[nodeid
]);
202 memset(NODE_DATA(nodeid
), 0, sizeof(pg_data_t
));
203 NODE_DATA(nodeid
)->bdata
= &plat_node_bdata
[nodeid
];
204 NODE_DATA(nodeid
)->node_start_pfn
= start_pfn
;
205 NODE_DATA(nodeid
)->node_spanned_pages
= end_pfn
- start_pfn
;
207 /* Find a place for the bootmem map */
208 bootmap_pages
= bootmem_bootmap_pages(end_pfn
- start_pfn
);
209 bootmap_start
= round_up(nodedata_phys
+ pgdat_size
, PAGE_SIZE
);
210 bootmap
= early_node_mem(nodeid
, bootmap_start
, end
,
211 bootmap_pages
<<PAGE_SHIFT
);
212 if (bootmap
== NULL
) {
213 if (nodedata_phys
< start
|| nodedata_phys
>= end
)
214 free_bootmem((unsigned long)node_data
[nodeid
],
216 node_data
[nodeid
] = NULL
;
219 bootmap_start
= __pa(bootmap
);
220 Dprintk("bootmap start %lu pages %lu\n", bootmap_start
, bootmap_pages
);
222 bootmap_size
= init_bootmem_node(NODE_DATA(nodeid
),
223 bootmap_start
>> PAGE_SHIFT
,
226 free_bootmem_with_active_regions(nodeid
, end
);
228 reserve_bootmem_node(NODE_DATA(nodeid
), nodedata_phys
, pgdat_size
);
229 reserve_bootmem_node(NODE_DATA(nodeid
), bootmap_start
,
230 bootmap_pages
<<PAGE_SHIFT
);
231 #ifdef CONFIG_ACPI_NUMA
232 srat_reserve_add_area(nodeid
);
234 node_set_online(nodeid
);
237 #ifdef CONFIG_FLAT_NODE_MEM_MAP
238 /* Initialize final allocator for a zone */
239 static void __init
flat_setup_node_zones(int nodeid
)
241 unsigned long start_pfn
, end_pfn
, memmapsize
, limit
;
243 start_pfn
= node_start_pfn(nodeid
);
244 end_pfn
= node_end_pfn(nodeid
);
246 Dprintk(KERN_INFO
"Setting up memmap for node %d %lx-%lx\n",
247 nodeid
, start_pfn
, end_pfn
);
250 * Try to allocate mem_map at end to not fill up precious <4GB
253 memmapsize
= sizeof(struct page
) * (end_pfn
-start_pfn
);
254 limit
= end_pfn
<< PAGE_SHIFT
;
256 NODE_DATA(nodeid
)->node_mem_map
=
257 __alloc_bootmem_core(NODE_DATA(nodeid
)->bdata
,
258 memmapsize
, SMP_CACHE_BYTES
,
259 round_down(limit
- memmapsize
, PAGE_SIZE
),
263 #define flat_setup_node_zones(i) do {} while (0)
267 * There are unfortunately some poorly designed mainboards around that
268 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
269 * mapping. To avoid this fill in the mapping for all possible CPUs,
270 * as the number of CPUs is not known yet. We round robin the existing
273 void __init
numa_init_array(void)
277 rr
= first_node(node_online_map
);
278 for (i
= 0; i
< NR_CPUS
; i
++) {
279 if (cpu_to_node(i
) != NUMA_NO_NODE
)
281 numa_set_node(i
, rr
);
282 rr
= next_node(rr
, node_online_map
);
283 if (rr
== MAX_NUMNODES
)
284 rr
= first_node(node_online_map
);
288 #ifdef CONFIG_NUMA_EMU
290 char *cmdline __initdata
;
293 * Setups up nid to range from addr to addr + size. If the end
294 * boundary is greater than max_addr, then max_addr is used instead.
295 * The return value is 0 if there is additional memory left for
296 * allocation past addr and -1 otherwise. addr is adjusted to be at
297 * the end of the node.
299 static int __init
setup_node_range(int nid
, struct bootnode
*nodes
, u64
*addr
,
300 u64 size
, u64 max_addr
)
304 nodes
[nid
].start
= *addr
;
306 if (*addr
>= max_addr
) {
310 nodes
[nid
].end
= *addr
;
311 node_set(nid
, node_possible_map
);
312 printk(KERN_INFO
"Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid
,
313 nodes
[nid
].start
, nodes
[nid
].end
,
314 (nodes
[nid
].end
- nodes
[nid
].start
) >> 20);
319 * Splits num_nodes nodes up equally starting at node_start. The return value
320 * is the number of nodes split up and addr is adjusted to be at the end of the
321 * last node allocated.
323 static int __init
split_nodes_equally(struct bootnode
*nodes
, u64
*addr
,
324 u64 max_addr
, int node_start
,
333 if (num_nodes
> MAX_NUMNODES
)
334 num_nodes
= MAX_NUMNODES
;
335 size
= (max_addr
- *addr
- e820_hole_size(*addr
, max_addr
)) /
338 * Calculate the number of big nodes that can be allocated as a result
339 * of consolidating the leftovers.
341 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * num_nodes
) /
344 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
345 size
&= FAKE_NODE_MIN_HASH_MASK
;
347 printk(KERN_ERR
"Not enough memory for each node. "
348 "NUMA emulation disabled.\n");
352 for (i
= node_start
; i
< num_nodes
+ node_start
; i
++) {
353 u64 end
= *addr
+ size
;
356 end
+= FAKE_NODE_MIN_SIZE
;
358 * The final node can have the remaining system RAM. Other
359 * nodes receive roughly the same amount of available pages.
361 if (i
== num_nodes
+ node_start
- 1)
364 while (end
- *addr
- e820_hole_size(*addr
, end
) <
366 end
+= FAKE_NODE_MIN_SIZE
;
367 if (end
> max_addr
) {
372 if (setup_node_range(i
, nodes
, addr
, end
- *addr
, max_addr
) < 0)
375 return i
- node_start
+ 1;
379 * Splits the remaining system RAM into chunks of size. The remaining memory is
380 * always assigned to a final node and can be asymmetric. Returns the number of
383 static int __init
split_nodes_by_size(struct bootnode
*nodes
, u64
*addr
,
384 u64 max_addr
, int node_start
, u64 size
)
387 size
= (size
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
388 while (!setup_node_range(i
++, nodes
, addr
, size
, max_addr
))
390 return i
- node_start
;
394 * Sets up the system RAM area from start_pfn to end_pfn according to the
395 * numa=fake command-line option.
397 static int __init
numa_emulation(unsigned long start_pfn
, unsigned long end_pfn
)
399 struct bootnode nodes
[MAX_NUMNODES
];
400 u64 size
, addr
= start_pfn
<< PAGE_SHIFT
;
401 u64 max_addr
= end_pfn
<< PAGE_SHIFT
;
402 int num_nodes
= 0, num
= 0, coeff_flag
, coeff
= -1, i
;
404 memset(&nodes
, 0, sizeof(nodes
));
406 * If the numa=fake command-line is just a single number N, split the
407 * system RAM into N fake nodes.
409 if (!strchr(cmdline
, '*') && !strchr(cmdline
, ',')) {
410 long n
= simple_strtol(cmdline
, NULL
, 0);
412 num_nodes
= split_nodes_equally(nodes
, &addr
, max_addr
, 0, n
);
418 /* Parse the command line. */
419 for (coeff_flag
= 0; ; cmdline
++) {
420 if (*cmdline
&& isdigit(*cmdline
)) {
421 num
= num
* 10 + *cmdline
- '0';
424 if (*cmdline
== '*') {
429 if (!*cmdline
|| *cmdline
== ',') {
433 * Round down to the nearest FAKE_NODE_MIN_SIZE.
434 * Command-line coefficients are in megabytes.
436 size
= ((u64
)num
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
438 for (i
= 0; i
< coeff
; i
++, num_nodes
++)
439 if (setup_node_range(num_nodes
, nodes
,
440 &addr
, size
, max_addr
) < 0)
452 /* Fill remainder of system RAM, if appropriate. */
453 if (addr
< max_addr
) {
454 if (coeff_flag
&& coeff
< 0) {
455 /* Split remaining nodes into num-sized chunks */
456 num_nodes
+= split_nodes_by_size(nodes
, &addr
, max_addr
,
460 switch (*(cmdline
- 1)) {
462 /* Split remaining nodes into coeff chunks */
465 num_nodes
+= split_nodes_equally(nodes
, &addr
, max_addr
,
469 /* Do not allocate remaining system RAM */
472 /* Give one final node */
473 setup_node_range(num_nodes
, nodes
, &addr
,
474 max_addr
- addr
, max_addr
);
479 memnode_shift
= compute_hash_shift(nodes
, num_nodes
);
480 if (memnode_shift
< 0) {
482 printk(KERN_ERR
"No NUMA hash function found. NUMA emulation "
488 * We need to vacate all active ranges that may have been registered by
489 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
490 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
492 remove_all_active_ranges();
493 #ifdef CONFIG_ACPI_NUMA
496 for_each_node_mask(i
, node_possible_map
) {
497 e820_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
498 nodes
[i
].end
>> PAGE_SHIFT
);
499 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
501 acpi_fake_nodes(nodes
, num_nodes
);
505 #endif /* CONFIG_NUMA_EMU */
507 void __init
numa_initmem_init(unsigned long start_pfn
, unsigned long end_pfn
)
511 nodes_clear(node_possible_map
);
513 #ifdef CONFIG_NUMA_EMU
514 if (cmdline
&& !numa_emulation(start_pfn
, end_pfn
))
516 nodes_clear(node_possible_map
);
519 #ifdef CONFIG_ACPI_NUMA
520 if (!numa_off
&& !acpi_scan_nodes(start_pfn
<< PAGE_SHIFT
,
521 end_pfn
<< PAGE_SHIFT
))
523 nodes_clear(node_possible_map
);
526 #ifdef CONFIG_K8_NUMA
527 if (!numa_off
&& !k8_scan_nodes(start_pfn
<<PAGE_SHIFT
,
528 end_pfn
<<PAGE_SHIFT
))
530 nodes_clear(node_possible_map
);
532 printk(KERN_INFO
"%s\n",
533 numa_off
? "NUMA turned off" : "No NUMA configuration found");
535 printk(KERN_INFO
"Faking a node at %016lx-%016lx\n",
536 start_pfn
<< PAGE_SHIFT
,
537 end_pfn
<< PAGE_SHIFT
);
538 /* setup dummy node covering all memory */
540 memnodemap
= memnode
.embedded_map
;
542 nodes_clear(node_online_map
);
544 node_set(0, node_possible_map
);
545 for (i
= 0; i
< NR_CPUS
; i
++)
547 /* we can't use cpumask_of_cpu() yet */
548 memset(&node_to_cpumask_map
[0], 0, sizeof(node_to_cpumask_map
[0]));
549 cpu_set(0, node_to_cpumask_map
[0]);
550 e820_register_active_regions(0, start_pfn
, end_pfn
);
551 setup_node_bootmem(0, start_pfn
<< PAGE_SHIFT
, end_pfn
<< PAGE_SHIFT
);
554 __cpuinit
void numa_add_cpu(int cpu
)
556 set_bit(cpu
, (unsigned long *)&node_to_cpumask_map
[cpu_to_node(cpu
)]);
559 void __cpuinit
numa_set_node(int cpu
, int node
)
561 cpu_pda(cpu
)->nodenumber
= node
;
562 cpu_to_node_map
[cpu
] = node
;
565 unsigned long __init
numa_free_all_bootmem(void)
567 unsigned long pages
= 0;
570 for_each_online_node(i
)
571 pages
+= free_all_bootmem_node(NODE_DATA(i
));
576 void __init
paging_init(void)
578 unsigned long max_zone_pfns
[MAX_NR_ZONES
];
581 memset(max_zone_pfns
, 0, sizeof(max_zone_pfns
));
582 max_zone_pfns
[ZONE_DMA
] = MAX_DMA_PFN
;
583 max_zone_pfns
[ZONE_DMA32
] = MAX_DMA32_PFN
;
584 max_zone_pfns
[ZONE_NORMAL
] = end_pfn
;
586 sparse_memory_present_with_active_regions(MAX_NUMNODES
);
589 for_each_online_node(i
)
590 flat_setup_node_zones(i
);
592 free_area_init_nodes(max_zone_pfns
);
595 static __init
int numa_setup(char *opt
)
599 if (!strncmp(opt
, "off", 3))
601 #ifdef CONFIG_NUMA_EMU
602 if (!strncmp(opt
, "fake=", 5))
605 #ifdef CONFIG_ACPI_NUMA
606 if (!strncmp(opt
, "noacpi", 6))
608 if (!strncmp(opt
, "hotadd=", 7))
609 hotadd_percent
= simple_strtoul(opt
+7, NULL
, 10);
613 early_param("numa", numa_setup
);
616 * Setup early cpu_to_node.
618 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
619 * and apicid_to_node[] tables have valid entries for a CPU.
620 * This means we skip cpu_to_node[] initialisation for NUMA
621 * emulation and faking node case (when running a kernel compiled
622 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
623 * is already initialized in a round robin manner at numa_init_array,
624 * prior to this call, and this initialization is good enough
625 * for the fake NUMA cases.
627 void __init
init_cpu_to_node(void)
631 for (i
= 0; i
< NR_CPUS
; i
++) {
632 u16 apicid
= x86_cpu_to_apicid_init
[i
];
634 if (apicid
== BAD_APICID
)
636 if (apicid_to_node
[apicid
] == NUMA_NO_NODE
)
638 numa_set_node(i
, apicid_to_node
[apicid
]);