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/memblock.h>
11 #include <linux/mmzone.h>
12 #include <linux/ctype.h>
13 #include <linux/module.h>
14 #include <linux/nodemask.h>
15 #include <linux/sched.h>
18 #include <asm/proto.h>
22 #include <asm/amd_nb.h>
24 struct pglist_data
*node_data
[MAX_NUMNODES
] __read_mostly
;
25 EXPORT_SYMBOL(node_data
);
27 struct memnode memnode
;
29 static unsigned long __initdata nodemap_addr
;
30 static unsigned long __initdata nodemap_size
;
33 * Given a shift value, try to populate memnodemap[]
36 * 0 if memnodmap[] too small (of shift too small)
37 * -1 if node overlap or lost ram (shift too big)
39 static int __init
populate_memnodemap(const struct bootnode
*nodes
,
40 int numnodes
, int shift
, int *nodeids
)
42 unsigned long addr
, end
;
45 memset(memnodemap
, 0xff, sizeof(s16
)*memnodemapsize
);
46 for (i
= 0; i
< numnodes
; i
++) {
47 addr
= nodes
[i
].start
;
51 if ((end
>> shift
) >= memnodemapsize
)
54 if (memnodemap
[addr
>> shift
] != NUMA_NO_NODE
)
58 memnodemap
[addr
>> shift
] = i
;
60 memnodemap
[addr
>> shift
] = nodeids
[i
];
62 addr
+= (1UL << shift
);
69 static int __init
allocate_cachealigned_memnodemap(void)
73 memnodemap
= memnode
.embedded_map
;
74 if (memnodemapsize
<= ARRAY_SIZE(memnode
.embedded_map
))
78 nodemap_size
= roundup(sizeof(s16
) * memnodemapsize
, L1_CACHE_BYTES
);
79 nodemap_addr
= memblock_find_in_range(addr
, get_max_mapped(),
80 nodemap_size
, L1_CACHE_BYTES
);
81 if (nodemap_addr
== MEMBLOCK_ERROR
) {
83 "NUMA: Unable to allocate Memory to Node hash map\n");
84 nodemap_addr
= nodemap_size
= 0;
87 memnodemap
= phys_to_virt(nodemap_addr
);
88 memblock_x86_reserve_range(nodemap_addr
, nodemap_addr
+ nodemap_size
, "MEMNODEMAP");
90 printk(KERN_DEBUG
"NUMA: Allocated memnodemap from %lx - %lx\n",
91 nodemap_addr
, nodemap_addr
+ nodemap_size
);
96 * The LSB of all start and end addresses in the node map is the value of the
97 * maximum possible shift.
99 static int __init
extract_lsb_from_nodes(const struct bootnode
*nodes
,
102 int i
, nodes_used
= 0;
103 unsigned long start
, end
;
104 unsigned long bitfield
= 0, memtop
= 0;
106 for (i
= 0; i
< numnodes
; i
++) {
107 start
= nodes
[i
].start
;
119 i
= find_first_bit(&bitfield
, sizeof(unsigned long)*8);
120 memnodemapsize
= (memtop
>> i
)+1;
124 int __init
compute_hash_shift(struct bootnode
*nodes
, int numnodes
,
129 shift
= extract_lsb_from_nodes(nodes
, numnodes
);
130 if (allocate_cachealigned_memnodemap())
132 printk(KERN_DEBUG
"NUMA: Using %d for the hash shift.\n",
135 if (populate_memnodemap(nodes
, numnodes
, shift
, nodeids
) != 1) {
136 printk(KERN_INFO
"Your memory is not aligned you need to "
137 "rebuild your kernel with a bigger NODEMAPSIZE "
138 "shift=%d\n", shift
);
144 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
146 return phys_to_nid(pfn
<< PAGE_SHIFT
);
149 static void * __init
early_node_mem(int nodeid
, unsigned long start
,
150 unsigned long end
, unsigned long size
,
156 * put it on high as possible
157 * something will go with NODE_DATA
159 if (start
< (MAX_DMA_PFN
<<PAGE_SHIFT
))
160 start
= MAX_DMA_PFN
<<PAGE_SHIFT
;
161 if (start
< (MAX_DMA32_PFN
<<PAGE_SHIFT
) &&
162 end
> (MAX_DMA32_PFN
<<PAGE_SHIFT
))
163 start
= MAX_DMA32_PFN
<<PAGE_SHIFT
;
164 mem
= memblock_x86_find_in_range_node(nodeid
, start
, end
, size
, align
);
165 if (mem
!= MEMBLOCK_ERROR
)
168 /* extend the search scope */
169 end
= max_pfn_mapped
<< PAGE_SHIFT
;
170 start
= MAX_DMA_PFN
<< PAGE_SHIFT
;
171 mem
= memblock_find_in_range(start
, end
, size
, align
);
172 if (mem
!= MEMBLOCK_ERROR
)
175 printk(KERN_ERR
"Cannot find %lu bytes in node %d\n",
181 /* Initialize bootmem allocator for a node */
183 setup_node_bootmem(int nodeid
, unsigned long start
, unsigned long end
)
185 unsigned long start_pfn
, last_pfn
, nodedata_phys
;
186 const int pgdat_size
= roundup(sizeof(pg_data_t
), PAGE_SIZE
);
193 * Don't confuse VM with a node that doesn't have the
194 * minimum amount of memory:
196 if (end
&& (end
- start
) < NODE_MIN_SIZE
)
199 start
= roundup(start
, ZONE_ALIGN
);
201 printk(KERN_INFO
"Initmem setup node %d %016lx-%016lx\n", nodeid
,
204 start_pfn
= start
>> PAGE_SHIFT
;
205 last_pfn
= end
>> PAGE_SHIFT
;
207 node_data
[nodeid
] = early_node_mem(nodeid
, start
, end
, pgdat_size
,
209 if (node_data
[nodeid
] == NULL
)
211 nodedata_phys
= __pa(node_data
[nodeid
]);
212 memblock_x86_reserve_range(nodedata_phys
, nodedata_phys
+ pgdat_size
, "NODE_DATA");
213 printk(KERN_INFO
" NODE_DATA [%016lx - %016lx]\n", nodedata_phys
,
214 nodedata_phys
+ pgdat_size
- 1);
215 nid
= phys_to_nid(nodedata_phys
);
217 printk(KERN_INFO
" NODE_DATA(%d) on node %d\n", nodeid
, nid
);
219 memset(NODE_DATA(nodeid
), 0, sizeof(pg_data_t
));
220 NODE_DATA(nodeid
)->node_id
= nodeid
;
221 NODE_DATA(nodeid
)->node_start_pfn
= start_pfn
;
222 NODE_DATA(nodeid
)->node_spanned_pages
= last_pfn
- start_pfn
;
224 node_set_online(nodeid
);
227 #ifdef CONFIG_NUMA_EMU
229 static struct bootnode nodes
[MAX_NUMNODES
] __initdata
;
230 static struct bootnode physnodes
[MAX_NUMNODES
] __cpuinitdata
;
231 static char *cmdline __initdata
;
233 void __init
numa_emu_cmdline(char *str
)
238 static int __init
setup_physnodes(unsigned long start
, unsigned long end
,
244 memset(physnodes
, 0, sizeof(physnodes
));
245 #ifdef CONFIG_ACPI_NUMA
247 acpi_get_nodes(physnodes
, start
, end
);
249 #ifdef CONFIG_AMD_NUMA
251 amd_get_nodes(physnodes
);
254 * Basic sanity checking on the physical node map: there may be errors
255 * if the SRAT or AMD code incorrectly reported the topology or the mem=
256 * kernel parameter is used.
258 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
259 if (physnodes
[i
].start
== physnodes
[i
].end
)
261 if (physnodes
[i
].start
> end
) {
262 physnodes
[i
].end
= physnodes
[i
].start
;
265 if (physnodes
[i
].end
< start
) {
266 physnodes
[i
].start
= physnodes
[i
].end
;
269 if (physnodes
[i
].start
< start
)
270 physnodes
[i
].start
= start
;
271 if (physnodes
[i
].end
> end
)
272 physnodes
[i
].end
= end
;
277 * If no physical topology was detected, a single node is faked to cover
278 * the entire address space.
281 physnodes
[ret
].start
= start
;
282 physnodes
[ret
].end
= end
;
288 static void __init
fake_physnodes(int acpi
, int amd
, int nr_nodes
)
293 #ifdef CONFIG_ACPI_NUMA
295 acpi_fake_nodes(nodes
, nr_nodes
);
297 #ifdef CONFIG_AMD_NUMA
299 amd_fake_nodes(nodes
, nr_nodes
);
302 for (i
= 0; i
< nr_cpu_ids
; i
++)
307 * Setups up nid to range from addr to addr + size. If the end
308 * boundary is greater than max_addr, then max_addr is used instead.
309 * The return value is 0 if there is additional memory left for
310 * allocation past addr and -1 otherwise. addr is adjusted to be at
311 * the end of the node.
313 static int __init
setup_node_range(int nid
, u64
*addr
, u64 size
, u64 max_addr
)
316 nodes
[nid
].start
= *addr
;
318 if (*addr
>= max_addr
) {
322 nodes
[nid
].end
= *addr
;
323 node_set(nid
, node_possible_map
);
324 printk(KERN_INFO
"Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid
,
325 nodes
[nid
].start
, nodes
[nid
].end
,
326 (nodes
[nid
].end
- nodes
[nid
].start
) >> 20);
331 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
332 * to max_addr. The return value is the number of nodes allocated.
334 static int __init
split_nodes_interleave(u64 addr
, u64 max_addr
, int nr_nodes
)
336 nodemask_t physnode_mask
= NODE_MASK_NONE
;
344 if (nr_nodes
> MAX_NUMNODES
) {
345 pr_info("numa=fake=%d too large, reducing to %d\n",
346 nr_nodes
, MAX_NUMNODES
);
347 nr_nodes
= MAX_NUMNODES
;
350 size
= (max_addr
- addr
- memblock_x86_hole_size(addr
, max_addr
)) / nr_nodes
;
352 * Calculate the number of big nodes that can be allocated as a result
353 * of consolidating the remainder.
355 big
= ((size
& ~FAKE_NODE_MIN_HASH_MASK
) * nr_nodes
) /
358 size
&= FAKE_NODE_MIN_HASH_MASK
;
360 pr_err("Not enough memory for each node. "
361 "NUMA emulation disabled.\n");
365 for (i
= 0; i
< MAX_NUMNODES
; i
++)
366 if (physnodes
[i
].start
!= physnodes
[i
].end
)
367 node_set(i
, physnode_mask
);
370 * Continue to fill physical nodes with fake nodes until there is no
371 * memory left on any of them.
373 while (nodes_weight(physnode_mask
)) {
374 for_each_node_mask(i
, physnode_mask
) {
375 u64 end
= physnodes
[i
].start
+ size
;
376 u64 dma32_end
= PFN_PHYS(MAX_DMA32_PFN
);
379 end
+= FAKE_NODE_MIN_SIZE
;
382 * Continue to add memory to this fake node if its
383 * non-reserved memory is less than the per-node size.
385 while (end
- physnodes
[i
].start
-
386 memblock_x86_hole_size(physnodes
[i
].start
, end
) < size
) {
387 end
+= FAKE_NODE_MIN_SIZE
;
388 if (end
> physnodes
[i
].end
) {
389 end
= physnodes
[i
].end
;
395 * If there won't be at least FAKE_NODE_MIN_SIZE of
396 * non-reserved memory in ZONE_DMA32 for the next node,
397 * this one must extend to the boundary.
399 if (end
< dma32_end
&& dma32_end
- end
-
400 memblock_x86_hole_size(end
, dma32_end
) < FAKE_NODE_MIN_SIZE
)
404 * If there won't be enough non-reserved memory for the
405 * next node, this one must extend to the end of the
408 if (physnodes
[i
].end
- end
-
409 memblock_x86_hole_size(end
, physnodes
[i
].end
) < size
)
410 end
= physnodes
[i
].end
;
413 * Avoid allocating more nodes than requested, which can
414 * happen as a result of rounding down each node's size
415 * to FAKE_NODE_MIN_SIZE.
417 if (nodes_weight(physnode_mask
) + ret
>= nr_nodes
)
418 end
= physnodes
[i
].end
;
420 if (setup_node_range(ret
++, &physnodes
[i
].start
,
421 end
- physnodes
[i
].start
,
422 physnodes
[i
].end
) < 0)
423 node_clear(i
, physnode_mask
);
430 * Returns the end address of a node so that there is at least `size' amount of
431 * non-reserved memory or `max_addr' is reached.
433 static u64 __init
find_end_of_node(u64 start
, u64 max_addr
, u64 size
)
435 u64 end
= start
+ size
;
437 while (end
- start
- memblock_x86_hole_size(start
, end
) < size
) {
438 end
+= FAKE_NODE_MIN_SIZE
;
439 if (end
> max_addr
) {
448 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
449 * `addr' to `max_addr'. The return value is the number of nodes allocated.
451 static int __init
split_nodes_size_interleave(u64 addr
, u64 max_addr
, u64 size
)
453 nodemask_t physnode_mask
= NODE_MASK_NONE
;
461 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
462 * increased accordingly if the requested size is too small. This
463 * creates a uniform distribution of node sizes across the entire
464 * machine (but not necessarily over physical nodes).
466 min_size
= (max_addr
- addr
- memblock_x86_hole_size(addr
, max_addr
)) /
468 min_size
= max(min_size
, FAKE_NODE_MIN_SIZE
);
469 if ((min_size
& FAKE_NODE_MIN_HASH_MASK
) < min_size
)
470 min_size
= (min_size
+ FAKE_NODE_MIN_SIZE
) &
471 FAKE_NODE_MIN_HASH_MASK
;
472 if (size
< min_size
) {
473 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
474 size
>> 20, min_size
>> 20);
477 size
&= FAKE_NODE_MIN_HASH_MASK
;
479 for (i
= 0; i
< MAX_NUMNODES
; i
++)
480 if (physnodes
[i
].start
!= physnodes
[i
].end
)
481 node_set(i
, physnode_mask
);
483 * Fill physical nodes with fake nodes of size until there is no memory
484 * left on any of them.
486 while (nodes_weight(physnode_mask
)) {
487 for_each_node_mask(i
, physnode_mask
) {
488 u64 dma32_end
= MAX_DMA32_PFN
<< PAGE_SHIFT
;
491 end
= find_end_of_node(physnodes
[i
].start
,
492 physnodes
[i
].end
, size
);
494 * If there won't be at least FAKE_NODE_MIN_SIZE of
495 * non-reserved memory in ZONE_DMA32 for the next node,
496 * this one must extend to the boundary.
498 if (end
< dma32_end
&& dma32_end
- end
-
499 memblock_x86_hole_size(end
, dma32_end
) < FAKE_NODE_MIN_SIZE
)
503 * If there won't be enough non-reserved memory for the
504 * next node, this one must extend to the end of the
507 if (physnodes
[i
].end
- end
-
508 memblock_x86_hole_size(end
, physnodes
[i
].end
) < size
)
509 end
= physnodes
[i
].end
;
512 * Setup the fake node that will be allocated as bootmem
513 * later. If setup_node_range() returns non-zero, there
514 * is no more memory available on this physical node.
516 if (setup_node_range(ret
++, &physnodes
[i
].start
,
517 end
- physnodes
[i
].start
,
518 physnodes
[i
].end
) < 0)
519 node_clear(i
, physnode_mask
);
526 * Sets up the system RAM area from start_pfn to last_pfn according to the
527 * numa=fake command-line option.
529 static int __init
numa_emulation(unsigned long start_pfn
,
530 unsigned long last_pfn
, int acpi
, int amd
)
532 u64 addr
= start_pfn
<< PAGE_SHIFT
;
533 u64 max_addr
= last_pfn
<< PAGE_SHIFT
;
538 * If the numa=fake command-line contains a 'M' or 'G', it represents
539 * the fixed node size. Otherwise, if it is just a single number N,
540 * split the system RAM into N fake nodes.
542 if (strchr(cmdline
, 'M') || strchr(cmdline
, 'G')) {
545 size
= memparse(cmdline
, &cmdline
);
546 num_nodes
= split_nodes_size_interleave(addr
, max_addr
, size
);
550 n
= simple_strtoul(cmdline
, NULL
, 0);
551 num_nodes
= split_nodes_interleave(addr
, max_addr
, n
);
556 memnode_shift
= compute_hash_shift(nodes
, num_nodes
, NULL
);
557 if (memnode_shift
< 0) {
559 printk(KERN_ERR
"No NUMA hash function found. NUMA emulation "
565 * We need to vacate all active ranges that may have been registered for
566 * the e820 memory map.
568 remove_all_active_ranges();
569 for_each_node_mask(i
, node_possible_map
)
570 memblock_x86_register_active_regions(i
, nodes
[i
].start
>> PAGE_SHIFT
,
571 nodes
[i
].end
>> PAGE_SHIFT
);
572 init_memory_mapping_high();
573 for_each_node_mask(i
, node_possible_map
)
574 setup_node_bootmem(i
, nodes
[i
].start
, nodes
[i
].end
);
575 setup_physnodes(addr
, max_addr
, acpi
, amd
);
576 fake_physnodes(acpi
, amd
, num_nodes
);
580 #endif /* CONFIG_NUMA_EMU */
582 void __init
initmem_init(int acpi
, int amd
)
586 nodes_clear(node_possible_map
);
587 nodes_clear(node_online_map
);
589 #ifdef CONFIG_NUMA_EMU
590 setup_physnodes(0, max_pfn
<< PAGE_SHIFT
, acpi
, amd
);
591 if (cmdline
&& !numa_emulation(0, max_pfn
, acpi
, amd
))
593 setup_physnodes(0, max_pfn
<< PAGE_SHIFT
, acpi
, amd
);
594 nodes_clear(node_possible_map
);
595 nodes_clear(node_online_map
);
598 #ifdef CONFIG_ACPI_NUMA
599 if (!numa_off
&& acpi
&& !acpi_scan_nodes())
601 nodes_clear(node_possible_map
);
602 nodes_clear(node_online_map
);
605 #ifdef CONFIG_AMD_NUMA
606 if (!numa_off
&& amd
&& !amd_scan_nodes())
608 nodes_clear(node_possible_map
);
609 nodes_clear(node_online_map
);
611 printk(KERN_INFO
"%s\n",
612 numa_off
? "NUMA turned off" : "No NUMA configuration found");
614 printk(KERN_INFO
"Faking a node at %016lx-%016lx\n",
615 0LU, max_pfn
<< PAGE_SHIFT
);
616 /* setup dummy node covering all memory */
618 memnodemap
= memnode
.embedded_map
;
621 node_set(0, node_possible_map
);
622 for (i
= 0; i
< MAX_LOCAL_APIC
; i
++)
623 set_apicid_to_node(i
, NUMA_NO_NODE
);
624 memblock_x86_register_active_regions(0, 0, max_pfn
);
625 init_memory_mapping_high();
626 setup_node_bootmem(0, 0, max_pfn
<< PAGE_SHIFT
);
630 unsigned long __init
numa_free_all_bootmem(void)
632 unsigned long pages
= 0;
635 for_each_online_node(i
)
636 pages
+= free_all_bootmem_node(NODE_DATA(i
));
638 pages
+= free_all_memory_core_early(MAX_NUMNODES
);
643 int __cpuinit
numa_cpu_node(int cpu
)
645 int apicid
= early_per_cpu(x86_cpu_to_apicid
, cpu
);
647 if (apicid
!= BAD_APICID
)
648 return __apicid_to_node
[apicid
];
653 * UGLINESS AHEAD: Currently, CONFIG_NUMA_EMU is 64bit only and makes use
654 * of 64bit specific data structures. The distinction is artificial and
655 * should be removed. numa_{add|remove}_cpu() are implemented in numa.c
656 * for both 32 and 64bit when CONFIG_NUMA_EMU is disabled but here when
659 * NUMA emulation is planned to be made generic and the following and other
660 * related code should be moved to numa.c.
662 #ifdef CONFIG_NUMA_EMU
663 # ifndef CONFIG_DEBUG_PER_CPU_MAPS
664 void __cpuinit
numa_add_cpu(int cpu
)
669 nid
= numa_cpu_node(cpu
);
670 if (nid
== NUMA_NO_NODE
)
671 nid
= early_cpu_to_node(cpu
);
672 BUG_ON(nid
== NUMA_NO_NODE
|| !node_online(nid
));
675 * Use the starting address of the emulated node to find which physical
676 * node it is allocated on.
678 addr
= node_start_pfn(nid
) << PAGE_SHIFT
;
679 for (physnid
= 0; physnid
< MAX_NUMNODES
; physnid
++)
680 if (addr
>= physnodes
[physnid
].start
&&
681 addr
< physnodes
[physnid
].end
)
685 * Map the cpu to each emulated node that is allocated on the physical
686 * node of the cpu's apic id.
688 for_each_online_node(nid
) {
689 addr
= node_start_pfn(nid
) << PAGE_SHIFT
;
690 if (addr
>= physnodes
[physnid
].start
&&
691 addr
< physnodes
[physnid
].end
)
692 cpumask_set_cpu(cpu
, node_to_cpumask_map
[nid
]);
696 void __cpuinit
numa_remove_cpu(int cpu
)
700 for_each_online_node(i
)
701 cpumask_clear_cpu(cpu
, node_to_cpumask_map
[i
]);
703 # else /* !CONFIG_DEBUG_PER_CPU_MAPS */
704 static void __cpuinit
numa_set_cpumask(int cpu
, int enable
)
706 int node
= early_cpu_to_node(cpu
);
707 struct cpumask
*mask
;
710 if (node
== NUMA_NO_NODE
) {
711 /* early_cpu_to_node() already emits a warning and trace */
714 for_each_online_node(i
) {
717 addr
= node_start_pfn(i
) << PAGE_SHIFT
;
718 if (addr
< physnodes
[node
].start
||
719 addr
>= physnodes
[node
].end
)
721 mask
= debug_cpumask_set_cpu(cpu
, enable
);
726 cpumask_set_cpu(cpu
, mask
);
728 cpumask_clear_cpu(cpu
, mask
);
732 void __cpuinit
numa_add_cpu(int cpu
)
734 numa_set_cpumask(cpu
, 1);
737 void __cpuinit
numa_remove_cpu(int cpu
)
739 numa_set_cpumask(cpu
, 0);
741 # endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
742 #endif /* CONFIG_NUMA_EMU */