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>
23 struct pglist_data
*node_data
[MAX_NUMNODES
] __read_mostly
;
24 EXPORT_SYMBOL(node_data
);
26 struct memnode memnode
;
28 s16 apicid_to_node
[MAX_LOCAL_APIC
] __cpuinitdata
= {
29 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
32 int numa_off __initdata
;
33 static unsigned long __initdata nodemap_addr
;
34 static unsigned long __initdata nodemap_size
;
37 * Given a shift value, try to populate memnodemap[]
40 * 0 if memnodmap[] too small (of shift too small)
41 * -1 if node overlap or lost ram (shift too big)
43 static int __init
populate_memnodemap(const struct bootnode
*nodes
,
44 int numnodes
, int shift
, int *nodeids
)
46 unsigned long addr
, end
;
49 memset(memnodemap
, 0xff, sizeof(s16
)*memnodemapsize
);
50 for (i
= 0; i
< numnodes
; i
++) {
51 addr
= nodes
[i
].start
;
55 if ((end
>> shift
) >= memnodemapsize
)
58 if (memnodemap
[addr
>> shift
] != NUMA_NO_NODE
)
62 memnodemap
[addr
>> shift
] = i
;
64 memnodemap
[addr
>> shift
] = nodeids
[i
];
66 addr
+= (1UL << shift
);
73 static int __init
allocate_cachealigned_memnodemap(void)
77 memnodemap
= memnode
.embedded_map
;
78 if (memnodemapsize
<= ARRAY_SIZE(memnode
.embedded_map
))
82 nodemap_size
= roundup(sizeof(s16
) * memnodemapsize
, L1_CACHE_BYTES
);
83 nodemap_addr
= find_e820_area(addr
, max_pfn
<<PAGE_SHIFT
,
84 nodemap_size
, L1_CACHE_BYTES
);
85 if (nodemap_addr
== -1UL) {
87 "NUMA: Unable to allocate Memory to Node hash map\n");
88 nodemap_addr
= nodemap_size
= 0;
91 memnodemap
= phys_to_virt(nodemap_addr
);
92 reserve_early(nodemap_addr
, nodemap_addr
+ nodemap_size
, "MEMNODEMAP");
94 printk(KERN_DEBUG
"NUMA: Allocated memnodemap from %lx - %lx\n",
95 nodemap_addr
, nodemap_addr
+ nodemap_size
);
100 * The LSB of all start and end addresses in the node map is the value of the
101 * maximum possible shift.
103 static int __init
extract_lsb_from_nodes(const struct bootnode
*nodes
,
106 int i
, nodes_used
= 0;
107 unsigned long start
, end
;
108 unsigned long bitfield
= 0, memtop
= 0;
110 for (i
= 0; i
< numnodes
; i
++) {
111 start
= nodes
[i
].start
;
123 i
= find_first_bit(&bitfield
, sizeof(unsigned long)*8);
124 memnodemapsize
= (memtop
>> i
)+1;
128 int __init
compute_hash_shift(struct bootnode
*nodes
, int numnodes
,
133 shift
= extract_lsb_from_nodes(nodes
, numnodes
);
134 if (allocate_cachealigned_memnodemap())
136 printk(KERN_DEBUG
"NUMA: Using %d for the hash shift.\n",
139 if (populate_memnodemap(nodes
, numnodes
, shift
, nodeids
) != 1) {
140 printk(KERN_INFO
"Your memory is not aligned you need to "
141 "rebuild your kernel with a bigger NODEMAPSIZE "
142 "shift=%d\n", shift
);
148 int early_pfn_to_nid(unsigned long pfn
)
150 return phys_to_nid(pfn
<< PAGE_SHIFT
);
153 static void * __init
early_node_mem(int nodeid
, unsigned long start
,
154 unsigned long end
, unsigned long size
,
157 unsigned long mem
= find_e820_area(start
, end
, size
, align
);
163 ptr
= __alloc_bootmem_nopanic(size
, align
, __pa(MAX_DMA_ADDRESS
));
165 printk(KERN_ERR
"Cannot find %lu bytes in node %d\n",
172 /* Initialize bootmem allocator for a node */
173 void __init
setup_node_bootmem(int nodeid
, unsigned long start
,
176 unsigned long start_pfn
, last_pfn
, bootmap_pages
, bootmap_size
;
177 unsigned long bootmap_start
, nodedata_phys
;
179 const int pgdat_size
= roundup(sizeof(pg_data_t
), PAGE_SIZE
);
182 start
= roundup(start
, ZONE_ALIGN
);
184 printk(KERN_INFO
"Bootmem setup node %d %016lx-%016lx\n", nodeid
,
187 start_pfn
= start
>> PAGE_SHIFT
;
188 last_pfn
= end
>> PAGE_SHIFT
;
190 node_data
[nodeid
] = early_node_mem(nodeid
, start
, end
, pgdat_size
,
192 if (node_data
[nodeid
] == NULL
)
194 nodedata_phys
= __pa(node_data
[nodeid
]);
195 printk(KERN_INFO
" NODE_DATA [%016lx - %016lx]\n", nodedata_phys
,
196 nodedata_phys
+ pgdat_size
- 1);
198 memset(NODE_DATA(nodeid
), 0, sizeof(pg_data_t
));
199 NODE_DATA(nodeid
)->bdata
= &bootmem_node_data
[nodeid
];
200 NODE_DATA(nodeid
)->node_start_pfn
= start_pfn
;
201 NODE_DATA(nodeid
)->node_spanned_pages
= last_pfn
- start_pfn
;
204 * Find a place for the bootmem map
205 * nodedata_phys could be on other nodes by alloc_bootmem,
206 * so need to sure bootmap_start not to be small, otherwise
207 * early_node_mem will get that with find_e820_area instead
208 * of alloc_bootmem, that could clash with reserved range
210 bootmap_pages
= bootmem_bootmap_pages(last_pfn
- start_pfn
);
211 nid
= phys_to_nid(nodedata_phys
);
213 bootmap_start
= roundup(nodedata_phys
+ pgdat_size
, PAGE_SIZE
);
215 bootmap_start
= roundup(start
, PAGE_SIZE
);
217 * SMP_CACHE_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
,
232 start_pfn
, last_pfn
);
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
);
241 * convert early reserve to bootmem reserve earlier
242 * otherwise early_node_mem could use early reserved mem
245 early_res_to_bootmem(start
, end
);
248 * in some case early_node_mem could use alloc_bootmem
249 * to get range on other node, don't reserve that again
252 printk(KERN_INFO
" NODE_DATA(%d) on node %d\n", nodeid
, nid
);
254 reserve_bootmem_node(NODE_DATA(nodeid
), nodedata_phys
,
255 pgdat_size
, BOOTMEM_DEFAULT
);
256 nid
= phys_to_nid(bootmap_start
);
258 printk(KERN_INFO
" bootmap(%d) on node %d\n", nodeid
, nid
);
260 reserve_bootmem_node(NODE_DATA(nodeid
), bootmap_start
,
261 bootmap_pages
<<PAGE_SHIFT
, BOOTMEM_DEFAULT
);
263 #ifdef CONFIG_ACPI_NUMA
264 srat_reserve_add_area(nodeid
);
266 node_set_online(nodeid
);
270 * There are unfortunately some poorly designed mainboards around that
271 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
272 * mapping. To avoid this fill in the mapping for all possible CPUs,
273 * as the number of CPUs is not known yet. We round robin the existing
276 void __init
numa_init_array(void)
280 rr
= first_node(node_online_map
);
281 for (i
= 0; i
< nr_cpu_ids
; i
++) {
282 if (early_cpu_to_node(i
) != NUMA_NO_NODE
)
284 numa_set_node(i
, rr
);
285 rr
= next_node(rr
, node_online_map
);
286 if (rr
== MAX_NUMNODES
)
287 rr
= first_node(node_online_map
);
291 #ifdef CONFIG_NUMA_EMU
293 static char *cmdline __initdata
;
296 * Setups up nid to range from addr to addr + size. If the end
297 * boundary is greater than max_addr, then max_addr is used instead.
298 * The return value is 0 if there is additional memory left for
299 * allocation past addr and -1 otherwise. addr is adjusted to be at
300 * the end of the node.
302 static int __init
setup_node_range(int nid
, struct bootnode
*nodes
, u64
*addr
,
303 u64 size
, u64 max_addr
)
307 nodes
[nid
].start
= *addr
;
309 if (*addr
>= max_addr
) {
313 nodes
[nid
].end
= *addr
;
314 node_set(nid
, node_possible_map
);
315 printk(KERN_INFO
"Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid
,
316 nodes
[nid
].start
, nodes
[nid
].end
,
317 (nodes
[nid
].end
- nodes
[nid
].start
) >> 20);
322 * Splits num_nodes nodes up equally starting at node_start. The return value
323 * is the number of nodes split up and addr is adjusted to be at the end of the
324 * last node allocated.
326 static int __init
split_nodes_equally(struct bootnode
*nodes
, u64
*addr
,
327 u64 max_addr
, int node_start
,
336 if (num_nodes
> MAX_NUMNODES
)
337 num_nodes
= MAX_NUMNODES
;
338 size
= (max_addr
- *addr
- e820_hole_size(*addr
, max_addr
)) /
341 * Calculate the number of big nodes that can be allocated as a result
342 * of consolidating the leftovers.
344 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * num_nodes
) /
347 /* Round down to nearest FAKE_NODE_MIN_SIZE. */
348 size
&= FAKE_NODE_MIN_HASH_MASK
;
350 printk(KERN_ERR
"Not enough memory for each node. "
351 "NUMA emulation disabled.\n");
355 for (i
= node_start
; i
< num_nodes
+ node_start
; i
++) {
356 u64 end
= *addr
+ size
;
359 end
+= FAKE_NODE_MIN_SIZE
;
361 * The final node can have the remaining system RAM. Other
362 * nodes receive roughly the same amount of available pages.
364 if (i
== num_nodes
+ node_start
- 1)
367 while (end
- *addr
- e820_hole_size(*addr
, end
) <
369 end
+= FAKE_NODE_MIN_SIZE
;
370 if (end
> max_addr
) {
375 if (setup_node_range(i
, nodes
, addr
, end
- *addr
, max_addr
) < 0)
378 return i
- node_start
+ 1;
382 * Splits the remaining system RAM into chunks of size. The remaining memory is
383 * always assigned to a final node and can be asymmetric. Returns the number of
386 static int __init
split_nodes_by_size(struct bootnode
*nodes
, u64
*addr
,
387 u64 max_addr
, int node_start
, u64 size
)
390 size
= (size
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
391 while (!setup_node_range(i
++, nodes
, addr
, size
, max_addr
))
393 return i
- node_start
;
397 * Sets up the system RAM area from start_pfn to last_pfn according to the
398 * numa=fake command-line option.
400 static struct bootnode nodes
[MAX_NUMNODES
] __initdata
;
402 static int __init
numa_emulation(unsigned long start_pfn
, unsigned long last_pfn
)
404 u64 size
, addr
= start_pfn
<< PAGE_SHIFT
;
405 u64 max_addr
= last_pfn
<< PAGE_SHIFT
;
406 int num_nodes
= 0, num
= 0, coeff_flag
, coeff
= -1, i
;
408 memset(&nodes
, 0, sizeof(nodes
));
410 * If the numa=fake command-line is just a single number N, split the
411 * system RAM into N fake nodes.
413 if (!strchr(cmdline
, '*') && !strchr(cmdline
, ',')) {
414 long n
= simple_strtol(cmdline
, NULL
, 0);
416 num_nodes
= split_nodes_equally(nodes
, &addr
, max_addr
, 0, n
);
422 /* Parse the command line. */
423 for (coeff_flag
= 0; ; cmdline
++) {
424 if (*cmdline
&& isdigit(*cmdline
)) {
425 num
= num
* 10 + *cmdline
- '0';
428 if (*cmdline
== '*') {
433 if (!*cmdline
|| *cmdline
== ',') {
437 * Round down to the nearest FAKE_NODE_MIN_SIZE.
438 * Command-line coefficients are in megabytes.
440 size
= ((u64
)num
<< 20) & FAKE_NODE_MIN_HASH_MASK
;
442 for (i
= 0; i
< coeff
; i
++, num_nodes
++)
443 if (setup_node_range(num_nodes
, nodes
,
444 &addr
, size
, max_addr
) < 0)
456 /* Fill remainder of system RAM, if appropriate. */
457 if (addr
< max_addr
) {
458 if (coeff_flag
&& coeff
< 0) {
459 /* Split remaining nodes into num-sized chunks */
460 num_nodes
+= split_nodes_by_size(nodes
, &addr
, max_addr
,
464 switch (*(cmdline
- 1)) {
466 /* Split remaining nodes into coeff chunks */
469 num_nodes
+= split_nodes_equally(nodes
, &addr
, max_addr
,
473 /* Do not allocate remaining system RAM */
476 /* Give one final node */
477 setup_node_range(num_nodes
, nodes
, &addr
,
478 max_addr
- addr
, max_addr
);
483 memnode_shift
= compute_hash_shift(nodes
, num_nodes
, NULL
);
484 if (memnode_shift
< 0) {
486 printk(KERN_ERR
"No NUMA hash function found. NUMA emulation "
492 * We need to vacate all active ranges that may have been registered by
493 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
494 * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
496 remove_all_active_ranges();
497 #ifdef CONFIG_ACPI_NUMA
500 for_each_node_mask(i
, node_possible_map
) {
501 e820_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
502 nodes
[i
].end
>> PAGE_SHIFT
);
503 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
505 acpi_fake_nodes(nodes
, num_nodes
);
509 #endif /* CONFIG_NUMA_EMU */
511 void __init
initmem_init(unsigned long start_pfn
, unsigned long last_pfn
)
515 nodes_clear(node_possible_map
);
516 nodes_clear(node_online_map
);
518 #ifdef CONFIG_NUMA_EMU
519 if (cmdline
&& !numa_emulation(start_pfn
, last_pfn
))
521 nodes_clear(node_possible_map
);
522 nodes_clear(node_online_map
);
525 #ifdef CONFIG_ACPI_NUMA
526 if (!numa_off
&& !acpi_scan_nodes(start_pfn
<< PAGE_SHIFT
,
527 last_pfn
<< PAGE_SHIFT
))
529 nodes_clear(node_possible_map
);
530 nodes_clear(node_online_map
);
533 #ifdef CONFIG_K8_NUMA
534 if (!numa_off
&& !k8_scan_nodes(start_pfn
<<PAGE_SHIFT
,
535 last_pfn
<<PAGE_SHIFT
))
537 nodes_clear(node_possible_map
);
538 nodes_clear(node_online_map
);
540 printk(KERN_INFO
"%s\n",
541 numa_off
? "NUMA turned off" : "No NUMA configuration found");
543 printk(KERN_INFO
"Faking a node at %016lx-%016lx\n",
544 start_pfn
<< PAGE_SHIFT
,
545 last_pfn
<< PAGE_SHIFT
);
546 /* setup dummy node covering all memory */
548 memnodemap
= memnode
.embedded_map
;
551 node_set(0, node_possible_map
);
552 for (i
= 0; i
< nr_cpu_ids
; i
++)
554 e820_register_active_regions(0, start_pfn
, last_pfn
);
555 setup_node_bootmem(0, start_pfn
<< PAGE_SHIFT
, last_pfn
<< PAGE_SHIFT
);
558 unsigned long __init
numa_free_all_bootmem(void)
560 unsigned long pages
= 0;
563 for_each_online_node(i
)
564 pages
+= free_all_bootmem_node(NODE_DATA(i
));
569 void __init
paging_init(void)
571 unsigned long max_zone_pfns
[MAX_NR_ZONES
];
573 memset(max_zone_pfns
, 0, sizeof(max_zone_pfns
));
574 max_zone_pfns
[ZONE_DMA
] = MAX_DMA_PFN
;
575 max_zone_pfns
[ZONE_DMA32
] = MAX_DMA32_PFN
;
576 max_zone_pfns
[ZONE_NORMAL
] = max_pfn
;
578 sparse_memory_present_with_active_regions(MAX_NUMNODES
);
581 free_area_init_nodes(max_zone_pfns
);
584 static __init
int numa_setup(char *opt
)
588 if (!strncmp(opt
, "off", 3))
590 #ifdef CONFIG_NUMA_EMU
591 if (!strncmp(opt
, "fake=", 5))
594 #ifdef CONFIG_ACPI_NUMA
595 if (!strncmp(opt
, "noacpi", 6))
597 if (!strncmp(opt
, "hotadd=", 7))
598 hotadd_percent
= simple_strtoul(opt
+7, NULL
, 10);
602 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 * Called before the per_cpu areas are setup.
619 void __init
init_cpu_to_node(void)
622 u16
*cpu_to_apicid
= early_per_cpu_ptr(x86_cpu_to_apicid
);
624 BUG_ON(cpu_to_apicid
== NULL
);
626 for_each_possible_cpu(cpu
) {
628 u16 apicid
= cpu_to_apicid
[cpu
];
630 if (apicid
== BAD_APICID
)
632 node
= apicid_to_node
[apicid
];
633 if (node
== NUMA_NO_NODE
)
635 if (!node_online(node
))
637 numa_set_node(cpu
, node
);