2 * Copyright (c) 2000, 2003 Silicon Graphics, Inc. All rights reserved.
3 * Copyright (c) 2001 Intel Corp.
4 * Copyright (c) 2001 Tony Luck <tony.luck@intel.com>
5 * Copyright (c) 2002 NEC Corp.
6 * Copyright (c) 2002 Kimio Suganuma <k-suganuma@da.jp.nec.com>
7 * Copyright (c) 2004 Silicon Graphics, Inc
8 * Russ Anderson <rja@sgi.com>
9 * Jesse Barnes <jbarnes@sgi.com>
10 * Jack Steiner <steiner@sgi.com>
14 * Platform initialization for Discontig Memory
17 #include <linux/kernel.h>
19 #include <linux/swap.h>
20 #include <linux/bootmem.h>
21 #include <linux/acpi.h>
22 #include <linux/efi.h>
23 #include <linux/nodemask.h>
24 #include <asm/pgalloc.h>
26 #include <asm/meminit.h>
28 #include <asm/sections.h>
31 * Track per-node information needed to setup the boot memory allocator, the
32 * per-node areas, and the real VM.
34 struct early_node_data
{
35 struct ia64_node_data
*node_data
;
37 unsigned long pernode_addr
;
38 unsigned long pernode_size
;
39 struct bootmem_data bootmem_data
;
40 unsigned long num_physpages
;
41 unsigned long num_dma_physpages
;
42 unsigned long min_pfn
;
43 unsigned long max_pfn
;
46 static struct early_node_data mem_data
[MAX_NUMNODES
] __initdata
;
47 static nodemask_t memory_less_mask __initdata
;
50 * To prevent cache aliasing effects, align per-node structures so that they
51 * start at addresses that are strided by node number.
53 #define MAX_NODE_ALIGN_OFFSET (32 * 1024 * 1024)
54 #define NODEDATA_ALIGN(addr, node) \
55 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + \
56 (((node)*PERCPU_PAGE_SIZE) & (MAX_NODE_ALIGN_OFFSET - 1)))
59 * build_node_maps - callback to setup bootmem structs for each node
60 * @start: physical start of range
61 * @len: length of range
62 * @node: node where this range resides
64 * We allocate a struct bootmem_data for each piece of memory that we wish to
65 * treat as a virtually contiguous block (i.e. each node). Each such block
66 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
67 * if necessary. Any non-existent pages will simply be part of the virtual
68 * memmap. We also update min_low_pfn and max_low_pfn here as we receive
69 * memory ranges from the caller.
71 static int __init
build_node_maps(unsigned long start
, unsigned long len
,
74 unsigned long cstart
, epfn
, end
= start
+ len
;
75 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
77 epfn
= GRANULEROUNDUP(end
) >> PAGE_SHIFT
;
78 cstart
= GRANULEROUNDDOWN(start
);
80 if (!bdp
->node_low_pfn
) {
81 bdp
->node_boot_start
= cstart
;
82 bdp
->node_low_pfn
= epfn
;
84 bdp
->node_boot_start
= min(cstart
, bdp
->node_boot_start
);
85 bdp
->node_low_pfn
= max(epfn
, bdp
->node_low_pfn
);
88 min_low_pfn
= min(min_low_pfn
, bdp
->node_boot_start
>>PAGE_SHIFT
);
89 max_low_pfn
= max(max_low_pfn
, bdp
->node_low_pfn
);
95 * early_nr_cpus_node - return number of cpus on a given node
96 * @node: node to check
98 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because
99 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
100 * called yet. Note that node 0 will also count all non-existent cpus.
102 static int __init
early_nr_cpus_node(int node
)
106 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
107 if (node
== node_cpuid
[cpu
].nid
)
114 * compute_pernodesize - compute size of pernode data
115 * @node: the node id.
117 static unsigned long __init
compute_pernodesize(int node
)
119 unsigned long pernodesize
= 0, cpus
;
121 cpus
= early_nr_cpus_node(node
);
122 pernodesize
+= PERCPU_PAGE_SIZE
* cpus
;
123 pernodesize
+= node
* L1_CACHE_BYTES
;
124 pernodesize
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
125 pernodesize
+= L1_CACHE_ALIGN(sizeof(struct ia64_node_data
));
126 pernodesize
= PAGE_ALIGN(pernodesize
);
131 * per_cpu_node_setup - setup per-cpu areas on each node
132 * @cpu_data: per-cpu area on this node
133 * @node: node to setup
135 * Copy the static per-cpu data into the region we just set aside and then
136 * setup __per_cpu_offset for each CPU on this node. Return a pointer to
137 * the end of the area.
139 static void *per_cpu_node_setup(void *cpu_data
, int node
)
144 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
145 if (node
== node_cpuid
[cpu
].nid
) {
146 memcpy(__va(cpu_data
), __phys_per_cpu_start
,
147 __per_cpu_end
- __per_cpu_start
);
148 __per_cpu_offset
[cpu
] = (char*)__va(cpu_data
) -
150 cpu_data
+= PERCPU_PAGE_SIZE
;
158 * fill_pernode - initialize pernode data.
159 * @node: the node id.
160 * @pernode: physical address of pernode data
161 * @pernodesize: size of the pernode data
163 static void __init
fill_pernode(int node
, unsigned long pernode
,
164 unsigned long pernodesize
)
167 int cpus
= early_nr_cpus_node(node
);
168 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
170 mem_data
[node
].pernode_addr
= pernode
;
171 mem_data
[node
].pernode_size
= pernodesize
;
172 memset(__va(pernode
), 0, pernodesize
);
174 cpu_data
= (void *)pernode
;
175 pernode
+= PERCPU_PAGE_SIZE
* cpus
;
176 pernode
+= node
* L1_CACHE_BYTES
;
178 mem_data
[node
].pgdat
= __va(pernode
);
179 pernode
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
181 mem_data
[node
].node_data
= __va(pernode
);
182 pernode
+= L1_CACHE_ALIGN(sizeof(struct ia64_node_data
));
184 mem_data
[node
].pgdat
->bdata
= bdp
;
185 pernode
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
187 cpu_data
= per_cpu_node_setup(cpu_data
, node
);
193 * find_pernode_space - allocate memory for memory map and per-node structures
194 * @start: physical start of range
195 * @len: length of range
196 * @node: node where this range resides
198 * This routine reserves space for the per-cpu data struct, the list of
199 * pg_data_ts and the per-node data struct. Each node will have something like
200 * the following in the first chunk of addr. space large enough to hold it.
202 * ________________________
204 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
205 * | PERCPU_PAGE_SIZE * | start and length big enough
206 * | cpus_on_this_node | Node 0 will also have entries for all non-existent cpus.
207 * |------------------------|
208 * | local pg_data_t * |
209 * |------------------------|
210 * | local ia64_node_data |
211 * |------------------------|
213 * |________________________|
215 * Once this space has been set aside, the bootmem maps are initialized. We
216 * could probably move the allocation of the per-cpu and ia64_node_data space
217 * outside of this function and use alloc_bootmem_node(), but doing it here
218 * is straightforward and we get the alignments we want so...
220 static int __init
find_pernode_space(unsigned long start
, unsigned long len
,
224 unsigned long pernodesize
= 0, pernode
, pages
, mapsize
;
225 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
227 epfn
= (start
+ len
) >> PAGE_SHIFT
;
229 pages
= bdp
->node_low_pfn
- (bdp
->node_boot_start
>> PAGE_SHIFT
);
230 mapsize
= bootmem_bootmap_pages(pages
) << PAGE_SHIFT
;
233 * Make sure this memory falls within this node's usable memory
234 * since we may have thrown some away in build_maps().
236 if (start
< bdp
->node_boot_start
|| epfn
> bdp
->node_low_pfn
)
239 /* Don't setup this node's local space twice... */
240 if (mem_data
[node
].pernode_addr
)
244 * Calculate total size needed, incl. what's necessary
245 * for good alignment and alias prevention.
247 pernodesize
= compute_pernodesize(node
);
248 pernode
= NODEDATA_ALIGN(start
, node
);
250 /* Is this range big enough for what we want to store here? */
251 if (start
+ len
> (pernode
+ pernodesize
+ mapsize
))
252 fill_pernode(node
, pernode
, pernodesize
);
258 * free_node_bootmem - free bootmem allocator memory for use
259 * @start: physical start of range
260 * @len: length of range
261 * @node: node where this range resides
263 * Simply calls the bootmem allocator to free the specified ranged from
264 * the given pg_data_t's bdata struct. After this function has been called
265 * for all the entries in the EFI memory map, the bootmem allocator will
266 * be ready to service allocation requests.
268 static int __init
free_node_bootmem(unsigned long start
, unsigned long len
,
271 free_bootmem_node(mem_data
[node
].pgdat
, start
, len
);
277 * reserve_pernode_space - reserve memory for per-node space
279 * Reserve the space used by the bootmem maps & per-node space in the boot
280 * allocator so that when we actually create the real mem maps we don't
283 static void __init
reserve_pernode_space(void)
285 unsigned long base
, size
, pages
;
286 struct bootmem_data
*bdp
;
289 for_each_online_node(node
) {
290 pg_data_t
*pdp
= mem_data
[node
].pgdat
;
292 if (node_isset(node
, memory_less_mask
))
297 /* First the bootmem_map itself */
298 pages
= bdp
->node_low_pfn
- (bdp
->node_boot_start
>>PAGE_SHIFT
);
299 size
= bootmem_bootmap_pages(pages
) << PAGE_SHIFT
;
300 base
= __pa(bdp
->node_bootmem_map
);
301 reserve_bootmem_node(pdp
, base
, size
);
303 /* Now the per-node space */
304 size
= mem_data
[node
].pernode_size
;
305 base
= __pa(mem_data
[node
].pernode_addr
);
306 reserve_bootmem_node(pdp
, base
, size
);
311 * initialize_pernode_data - fixup per-cpu & per-node pointers
313 * Each node's per-node area has a copy of the global pg_data_t list, so
314 * we copy that to each node here, as well as setting the per-cpu pointer
315 * to the local node data structure. The active_cpus field of the per-node
316 * structure gets setup by the platform_cpu_init() function later.
318 static void __init
initialize_pernode_data(void)
320 pg_data_t
*pgdat_list
[MAX_NUMNODES
];
323 for_each_online_node(node
)
324 pgdat_list
[node
] = mem_data
[node
].pgdat
;
326 /* Copy the pg_data_t list to each node and init the node field */
327 for_each_online_node(node
) {
328 memcpy(mem_data
[node
].node_data
->pg_data_ptrs
, pgdat_list
,
332 /* Set the node_data pointer for each per-cpu struct */
333 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
334 node
= node_cpuid
[cpu
].nid
;
335 per_cpu(cpu_info
, cpu
).node_data
= mem_data
[node
].node_data
;
339 struct cpuinfo_ia64
*cpu0_cpu_info
;
341 node
= node_cpuid
[cpu
].nid
;
342 cpu0_cpu_info
= (struct cpuinfo_ia64
*)(__phys_per_cpu_start
+
343 ((char *)&per_cpu__cpu_info
- __per_cpu_start
));
344 cpu0_cpu_info
->node_data
= mem_data
[node
].node_data
;
346 #endif /* CONFIG_SMP */
350 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit
351 * node but fall back to any other node when __alloc_bootmem_node fails
354 * @pernodesize: size of this node's pernode data
356 static void __init
*memory_less_node_alloc(int nid
, unsigned long pernodesize
)
360 int bestnode
= -1, node
, anynode
= 0;
362 for_each_online_node(node
) {
363 if (node_isset(node
, memory_less_mask
))
365 else if (node_distance(nid
, node
) < best
) {
366 best
= node_distance(nid
, node
);
375 ptr
= __alloc_bootmem_node(mem_data
[bestnode
].pgdat
, pernodesize
,
376 PERCPU_PAGE_SIZE
, __pa(MAX_DMA_ADDRESS
));
382 * pgdat_insert - insert the pgdat into global pgdat_list
383 * @pgdat: the pgdat for a node.
385 static void __init
pgdat_insert(pg_data_t
*pgdat
)
387 pg_data_t
*prev
= NULL
, *next
;
390 if (pgdat
->node_id
< next
->node_id
)
396 prev
->pgdat_next
= pgdat
;
397 pgdat
->pgdat_next
= next
;
399 pgdat
->pgdat_next
= pgdat_list
;
407 * memory_less_nodes - allocate and initialize CPU only nodes pernode
410 static void __init
memory_less_nodes(void)
412 unsigned long pernodesize
;
416 for_each_node_mask(node
, memory_less_mask
) {
417 pernodesize
= compute_pernodesize(node
);
418 pernode
= memory_less_node_alloc(node
, pernodesize
);
419 fill_pernode(node
, __pa(pernode
), pernodesize
);
425 #ifdef CONFIG_SPARSEMEM
427 * register_sparse_mem - notify SPARSEMEM that this memory range exists.
428 * @start: physical start of range
429 * @end: physical end of range
432 * Simply calls SPARSEMEM to register memory section(s).
434 static int __init
register_sparse_mem(unsigned long start
, unsigned long end
,
439 start
= __pa(start
) >> PAGE_SHIFT
;
440 end
= __pa(end
) >> PAGE_SHIFT
;
441 nid
= early_pfn_to_nid(start
);
442 memory_present(nid
, start
, end
);
447 static void __init
arch_sparse_init(void)
449 efi_memmap_walk(register_sparse_mem
, NULL
);
453 #define arch_sparse_init() do {} while (0)
457 * find_memory - walk the EFI memory map and setup the bootmem allocator
459 * Called early in boot to setup the bootmem allocator, and to
460 * allocate the per-cpu and per-node structures.
462 void __init
find_memory(void)
468 if (num_online_nodes() == 0) {
469 printk(KERN_ERR
"node info missing!\n");
473 nodes_or(memory_less_mask
, memory_less_mask
, node_online_map
);
477 /* These actually end up getting called by call_pernode_memory() */
478 efi_memmap_walk(filter_rsvd_memory
, build_node_maps
);
479 efi_memmap_walk(filter_rsvd_memory
, find_pernode_space
);
481 for_each_online_node(node
)
482 if (mem_data
[node
].bootmem_data
.node_low_pfn
) {
483 node_clear(node
, memory_less_mask
);
484 mem_data
[node
].min_pfn
= ~0UL;
487 * Initialize the boot memory maps in reverse order since that's
488 * what the bootmem allocator expects
490 for (node
= MAX_NUMNODES
- 1; node
>= 0; node
--) {
491 unsigned long pernode
, pernodesize
, map
;
492 struct bootmem_data
*bdp
;
494 if (!node_online(node
))
496 else if (node_isset(node
, memory_less_mask
))
499 bdp
= &mem_data
[node
].bootmem_data
;
500 pernode
= mem_data
[node
].pernode_addr
;
501 pernodesize
= mem_data
[node
].pernode_size
;
502 map
= pernode
+ pernodesize
;
504 init_bootmem_node(mem_data
[node
].pgdat
,
506 bdp
->node_boot_start
>>PAGE_SHIFT
,
510 efi_memmap_walk(filter_rsvd_memory
, free_node_bootmem
);
512 reserve_pernode_space();
514 initialize_pernode_data();
516 max_pfn
= max_low_pfn
;
523 * per_cpu_init - setup per-cpu variables
525 * find_pernode_space() does most of this already, we just need to set
526 * local_per_cpu_offset
528 void *per_cpu_init(void)
532 if (smp_processor_id() != 0)
533 return __per_cpu_start
+ __per_cpu_offset
[smp_processor_id()];
535 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
536 per_cpu(local_per_cpu_offset
, cpu
) = __per_cpu_offset
[cpu
];
538 return __per_cpu_start
+ __per_cpu_offset
[smp_processor_id()];
540 #endif /* CONFIG_SMP */
543 * show_mem - give short summary of memory stats
545 * Shows a simple page count of reserved and used pages in the system.
546 * For discontig machines, it does this on a per-pgdat basis.
550 int i
, total_reserved
= 0;
551 int total_shared
= 0, total_cached
= 0;
552 unsigned long total_present
= 0;
555 printk("Mem-info:\n");
557 printk("Free swap: %6ldkB\n", nr_swap_pages
<<(PAGE_SHIFT
-10));
558 for_each_pgdat(pgdat
) {
559 unsigned long present
;
561 int shared
= 0, cached
= 0, reserved
= 0;
563 printk("Node ID: %d\n", pgdat
->node_id
);
564 pgdat_resize_lock(pgdat
, &flags
);
565 present
= pgdat
->node_present_pages
;
566 for(i
= 0; i
< pgdat
->node_spanned_pages
; i
++) {
568 if (pfn_valid(pgdat
->node_start_pfn
+ i
))
569 page
= pfn_to_page(pgdat
->node_start_pfn
+ i
);
572 if (PageReserved(page
))
574 else if (PageSwapCache(page
))
576 else if (page_count(page
))
577 shared
+= page_count(page
)-1;
579 pgdat_resize_unlock(pgdat
, &flags
);
580 total_present
+= present
;
581 total_reserved
+= reserved
;
582 total_cached
+= cached
;
583 total_shared
+= shared
;
584 printk("\t%ld pages of RAM\n", present
);
585 printk("\t%d reserved pages\n", reserved
);
586 printk("\t%d pages shared\n", shared
);
587 printk("\t%d pages swap cached\n", cached
);
589 printk("%ld pages of RAM\n", total_present
);
590 printk("%d reserved pages\n", total_reserved
);
591 printk("%d pages shared\n", total_shared
);
592 printk("%d pages swap cached\n", total_cached
);
593 printk("Total of %ld pages in page table cache\n",
594 pgtable_quicklist_total_size());
595 printk("%d free buffer pages\n", nr_free_buffer_pages());
599 * call_pernode_memory - use SRAT to call callback functions with node info
600 * @start: physical start of range
601 * @len: length of range
602 * @arg: function to call for each range
604 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
605 * out to which node a block of memory belongs. Ignore memory that we cannot
606 * identify, and split blocks that run across multiple nodes.
608 * Take this opportunity to round the start address up and the end address
609 * down to page boundaries.
611 void call_pernode_memory(unsigned long start
, unsigned long len
, void *arg
)
613 unsigned long rs
, re
, end
= start
+ len
;
614 void (*func
)(unsigned long, unsigned long, int);
617 start
= PAGE_ALIGN(start
);
624 if (!num_node_memblks
) {
625 /* No SRAT table, so assume one node (node 0) */
627 (*func
)(start
, end
- start
, 0);
631 for (i
= 0; i
< num_node_memblks
; i
++) {
632 rs
= max(start
, node_memblk
[i
].start_paddr
);
633 re
= min(end
, node_memblk
[i
].start_paddr
+
634 node_memblk
[i
].size
);
637 (*func
)(rs
, re
- rs
, node_memblk
[i
].nid
);
645 * count_node_pages - callback to build per-node memory info structures
646 * @start: physical start of range
647 * @len: length of range
648 * @node: node where this range resides
650 * Each node has it's own number of physical pages, DMAable pages, start, and
651 * end page frame number. This routine will be called by call_pernode_memory()
652 * for each piece of usable memory and will setup these values for each node.
653 * Very similar to build_maps().
655 static __init
int count_node_pages(unsigned long start
, unsigned long len
, int node
)
657 unsigned long end
= start
+ len
;
659 mem_data
[node
].num_physpages
+= len
>> PAGE_SHIFT
;
660 if (start
<= __pa(MAX_DMA_ADDRESS
))
661 mem_data
[node
].num_dma_physpages
+=
662 (min(end
, __pa(MAX_DMA_ADDRESS
)) - start
) >>PAGE_SHIFT
;
663 start
= GRANULEROUNDDOWN(start
);
664 start
= ORDERROUNDDOWN(start
);
665 end
= GRANULEROUNDUP(end
);
666 mem_data
[node
].max_pfn
= max(mem_data
[node
].max_pfn
,
668 mem_data
[node
].min_pfn
= min(mem_data
[node
].min_pfn
,
669 start
>> PAGE_SHIFT
);
675 * paging_init - setup page tables
677 * paging_init() sets up the page tables for each node of the system and frees
678 * the bootmem allocator memory for general use.
680 void __init
paging_init(void)
682 unsigned long max_dma
;
683 unsigned long zones_size
[MAX_NR_ZONES
];
684 unsigned long zholes_size
[MAX_NR_ZONES
];
685 unsigned long pfn_offset
= 0;
688 max_dma
= virt_to_phys((void *) MAX_DMA_ADDRESS
) >> PAGE_SHIFT
;
692 efi_memmap_walk(filter_rsvd_memory
, count_node_pages
);
694 #ifdef CONFIG_VIRTUAL_MEM_MAP
695 vmalloc_end
-= PAGE_ALIGN(max_low_pfn
* sizeof(struct page
));
696 vmem_map
= (struct page
*) vmalloc_end
;
697 efi_memmap_walk(create_mem_map_page_table
, NULL
);
698 printk("Virtual mem_map starts at 0x%p\n", vmem_map
);
701 for_each_online_node(node
) {
702 memset(zones_size
, 0, sizeof(zones_size
));
703 memset(zholes_size
, 0, sizeof(zholes_size
));
705 num_physpages
+= mem_data
[node
].num_physpages
;
707 if (mem_data
[node
].min_pfn
>= max_dma
) {
708 /* All of this node's memory is above ZONE_DMA */
709 zones_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
710 mem_data
[node
].min_pfn
;
711 zholes_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
712 mem_data
[node
].min_pfn
-
713 mem_data
[node
].num_physpages
;
714 } else if (mem_data
[node
].max_pfn
< max_dma
) {
715 /* All of this node's memory is in ZONE_DMA */
716 zones_size
[ZONE_DMA
] = mem_data
[node
].max_pfn
-
717 mem_data
[node
].min_pfn
;
718 zholes_size
[ZONE_DMA
] = mem_data
[node
].max_pfn
-
719 mem_data
[node
].min_pfn
-
720 mem_data
[node
].num_dma_physpages
;
722 /* This node has memory in both zones */
723 zones_size
[ZONE_DMA
] = max_dma
-
724 mem_data
[node
].min_pfn
;
725 zholes_size
[ZONE_DMA
] = zones_size
[ZONE_DMA
] -
726 mem_data
[node
].num_dma_physpages
;
727 zones_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
729 zholes_size
[ZONE_NORMAL
] = zones_size
[ZONE_NORMAL
] -
730 (mem_data
[node
].num_physpages
-
731 mem_data
[node
].num_dma_physpages
);
734 pfn_offset
= mem_data
[node
].min_pfn
;
736 #ifdef CONFIG_VIRTUAL_MEM_MAP
737 NODE_DATA(node
)->node_mem_map
= vmem_map
+ pfn_offset
;
739 free_area_init_node(node
, NODE_DATA(node
), zones_size
,
740 pfn_offset
, zholes_size
);
744 * Make memory less nodes become a member of the known nodes.
746 for_each_node_mask(node
, memory_less_mask
)
747 pgdat_insert(mem_data
[node
].pgdat
);
749 zero_page_memmap_ptr
= virt_to_page(ia64_imva(empty_zero_page
));