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
;
36 unsigned long pernode_addr
;
37 unsigned long pernode_size
;
38 struct bootmem_data bootmem_data
;
39 unsigned long num_physpages
;
40 unsigned long num_dma_physpages
;
41 unsigned long min_pfn
;
42 unsigned long max_pfn
;
45 static struct early_node_data mem_data
[MAX_NUMNODES
] __initdata
;
46 static nodemask_t memory_less_mask __initdata
;
48 static pg_data_t
*pgdat_list
[MAX_NUMNODES
];
51 * To prevent cache aliasing effects, align per-node structures so that they
52 * start at addresses that are strided by node number.
54 #define MAX_NODE_ALIGN_OFFSET (32 * 1024 * 1024)
55 #define NODEDATA_ALIGN(addr, node) \
56 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + \
57 (((node)*PERCPU_PAGE_SIZE) & (MAX_NODE_ALIGN_OFFSET - 1)))
60 * build_node_maps - callback to setup bootmem structs for each node
61 * @start: physical start of range
62 * @len: length of range
63 * @node: node where this range resides
65 * We allocate a struct bootmem_data for each piece of memory that we wish to
66 * treat as a virtually contiguous block (i.e. each node). Each such block
67 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
68 * if necessary. Any non-existent pages will simply be part of the virtual
69 * memmap. We also update min_low_pfn and max_low_pfn here as we receive
70 * memory ranges from the caller.
72 static int __init
build_node_maps(unsigned long start
, unsigned long len
,
75 unsigned long cstart
, epfn
, end
= start
+ len
;
76 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
78 epfn
= GRANULEROUNDUP(end
) >> PAGE_SHIFT
;
79 cstart
= GRANULEROUNDDOWN(start
);
81 if (!bdp
->node_low_pfn
) {
82 bdp
->node_boot_start
= cstart
;
83 bdp
->node_low_pfn
= epfn
;
85 bdp
->node_boot_start
= min(cstart
, bdp
->node_boot_start
);
86 bdp
->node_low_pfn
= max(epfn
, bdp
->node_low_pfn
);
89 min_low_pfn
= min(min_low_pfn
, bdp
->node_boot_start
>>PAGE_SHIFT
);
90 max_low_pfn
= max(max_low_pfn
, bdp
->node_low_pfn
);
96 * early_nr_cpus_node - return number of cpus on a given node
97 * @node: node to check
99 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because
100 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
101 * called yet. Note that node 0 will also count all non-existent cpus.
103 static int __meminit
early_nr_cpus_node(int node
)
107 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
108 if (node
== node_cpuid
[cpu
].nid
)
115 * compute_pernodesize - compute size of pernode data
116 * @node: the node id.
118 static unsigned long __meminit
compute_pernodesize(int node
)
120 unsigned long pernodesize
= 0, cpus
;
122 cpus
= early_nr_cpus_node(node
);
123 pernodesize
+= PERCPU_PAGE_SIZE
* cpus
;
124 pernodesize
+= node
* L1_CACHE_BYTES
;
125 pernodesize
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
126 pernodesize
+= L1_CACHE_ALIGN(sizeof(struct ia64_node_data
));
127 pernodesize
= PAGE_ALIGN(pernodesize
);
132 * per_cpu_node_setup - setup per-cpu areas on each node
133 * @cpu_data: per-cpu area on this node
134 * @node: node to setup
136 * Copy the static per-cpu data into the region we just set aside and then
137 * setup __per_cpu_offset for each CPU on this node. Return a pointer to
138 * the end of the area.
140 static void *per_cpu_node_setup(void *cpu_data
, int node
)
145 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
146 if (node
== node_cpuid
[cpu
].nid
) {
147 memcpy(__va(cpu_data
), __phys_per_cpu_start
,
148 __per_cpu_end
- __per_cpu_start
);
149 __per_cpu_offset
[cpu
] = (char*)__va(cpu_data
) -
151 cpu_data
+= PERCPU_PAGE_SIZE
;
159 * fill_pernode - initialize pernode data.
160 * @node: the node id.
161 * @pernode: physical address of pernode data
162 * @pernodesize: size of the pernode data
164 static void __init
fill_pernode(int node
, unsigned long pernode
,
165 unsigned long pernodesize
)
168 int cpus
= early_nr_cpus_node(node
);
169 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
171 mem_data
[node
].pernode_addr
= pernode
;
172 mem_data
[node
].pernode_size
= pernodesize
;
173 memset(__va(pernode
), 0, pernodesize
);
175 cpu_data
= (void *)pernode
;
176 pernode
+= PERCPU_PAGE_SIZE
* cpus
;
177 pernode
+= node
* L1_CACHE_BYTES
;
179 pgdat_list
[node
] = __va(pernode
);
180 pernode
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
182 mem_data
[node
].node_data
= __va(pernode
);
183 pernode
+= L1_CACHE_ALIGN(sizeof(struct ia64_node_data
));
185 pgdat_list
[node
]->bdata
= bdp
;
186 pernode
+= L1_CACHE_ALIGN(sizeof(pg_data_t
));
188 cpu_data
= per_cpu_node_setup(cpu_data
, node
);
194 * find_pernode_space - allocate memory for memory map and per-node structures
195 * @start: physical start of range
196 * @len: length of range
197 * @node: node where this range resides
199 * This routine reserves space for the per-cpu data struct, the list of
200 * pg_data_ts and the per-node data struct. Each node will have something like
201 * the following in the first chunk of addr. space large enough to hold it.
203 * ________________________
205 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
206 * | PERCPU_PAGE_SIZE * | start and length big enough
207 * | cpus_on_this_node | Node 0 will also have entries for all non-existent cpus.
208 * |------------------------|
209 * | local pg_data_t * |
210 * |------------------------|
211 * | local ia64_node_data |
212 * |------------------------|
214 * |________________________|
216 * Once this space has been set aside, the bootmem maps are initialized. We
217 * could probably move the allocation of the per-cpu and ia64_node_data space
218 * outside of this function and use alloc_bootmem_node(), but doing it here
219 * is straightforward and we get the alignments we want so...
221 static int __init
find_pernode_space(unsigned long start
, unsigned long len
,
225 unsigned long pernodesize
= 0, pernode
, pages
, mapsize
;
226 struct bootmem_data
*bdp
= &mem_data
[node
].bootmem_data
;
228 epfn
= (start
+ len
) >> PAGE_SHIFT
;
230 pages
= bdp
->node_low_pfn
- (bdp
->node_boot_start
>> PAGE_SHIFT
);
231 mapsize
= bootmem_bootmap_pages(pages
) << PAGE_SHIFT
;
234 * Make sure this memory falls within this node's usable memory
235 * since we may have thrown some away in build_maps().
237 if (start
< bdp
->node_boot_start
|| epfn
> bdp
->node_low_pfn
)
240 /* Don't setup this node's local space twice... */
241 if (mem_data
[node
].pernode_addr
)
245 * Calculate total size needed, incl. what's necessary
246 * for good alignment and alias prevention.
248 pernodesize
= compute_pernodesize(node
);
249 pernode
= NODEDATA_ALIGN(start
, node
);
251 /* Is this range big enough for what we want to store here? */
252 if (start
+ len
> (pernode
+ pernodesize
+ mapsize
))
253 fill_pernode(node
, pernode
, pernodesize
);
259 * free_node_bootmem - free bootmem allocator memory for use
260 * @start: physical start of range
261 * @len: length of range
262 * @node: node where this range resides
264 * Simply calls the bootmem allocator to free the specified ranged from
265 * the given pg_data_t's bdata struct. After this function has been called
266 * for all the entries in the EFI memory map, the bootmem allocator will
267 * be ready to service allocation requests.
269 static int __init
free_node_bootmem(unsigned long start
, unsigned long len
,
272 free_bootmem_node(pgdat_list
[node
], start
, len
);
278 * reserve_pernode_space - reserve memory for per-node space
280 * Reserve the space used by the bootmem maps & per-node space in the boot
281 * allocator so that when we actually create the real mem maps we don't
284 static void __init
reserve_pernode_space(void)
286 unsigned long base
, size
, pages
;
287 struct bootmem_data
*bdp
;
290 for_each_online_node(node
) {
291 pg_data_t
*pdp
= pgdat_list
[node
];
293 if (node_isset(node
, memory_less_mask
))
298 /* First the bootmem_map itself */
299 pages
= bdp
->node_low_pfn
- (bdp
->node_boot_start
>>PAGE_SHIFT
);
300 size
= bootmem_bootmap_pages(pages
) << PAGE_SHIFT
;
301 base
= __pa(bdp
->node_bootmem_map
);
302 reserve_bootmem_node(pdp
, base
, size
);
304 /* Now the per-node space */
305 size
= mem_data
[node
].pernode_size
;
306 base
= __pa(mem_data
[node
].pernode_addr
);
307 reserve_bootmem_node(pdp
, base
, size
);
311 static void __meminit
scatter_node_data(void)
317 * for_each_online_node() can't be used at here.
318 * node_online_map is not set for hot-added nodes at this time,
319 * because we are halfway through initialization of the new node's
320 * structures. If for_each_online_node() is used, a new node's
321 * pg_data_ptrs will be not initialized. Insted of using it,
322 * pgdat_list[] is checked.
324 for_each_node(node
) {
325 if (pgdat_list
[node
]) {
326 dst
= LOCAL_DATA_ADDR(pgdat_list
[node
])->pg_data_ptrs
;
327 memcpy(dst
, pgdat_list
, sizeof(pgdat_list
));
333 * initialize_pernode_data - fixup per-cpu & per-node pointers
335 * Each node's per-node area has a copy of the global pg_data_t list, so
336 * we copy that to each node here, as well as setting the per-cpu pointer
337 * to the local node data structure. The active_cpus field of the per-node
338 * structure gets setup by the platform_cpu_init() function later.
340 static void __init
initialize_pernode_data(void)
347 /* Set the node_data pointer for each per-cpu struct */
348 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
349 node
= node_cpuid
[cpu
].nid
;
350 per_cpu(cpu_info
, cpu
).node_data
= mem_data
[node
].node_data
;
354 struct cpuinfo_ia64
*cpu0_cpu_info
;
356 node
= node_cpuid
[cpu
].nid
;
357 cpu0_cpu_info
= (struct cpuinfo_ia64
*)(__phys_per_cpu_start
+
358 ((char *)&per_cpu__cpu_info
- __per_cpu_start
));
359 cpu0_cpu_info
->node_data
= mem_data
[node
].node_data
;
361 #endif /* CONFIG_SMP */
365 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit
366 * node but fall back to any other node when __alloc_bootmem_node fails
369 * @pernodesize: size of this node's pernode data
371 static void __init
*memory_less_node_alloc(int nid
, unsigned long pernodesize
)
375 int bestnode
= -1, node
, anynode
= 0;
377 for_each_online_node(node
) {
378 if (node_isset(node
, memory_less_mask
))
380 else if (node_distance(nid
, node
) < best
) {
381 best
= node_distance(nid
, node
);
390 ptr
= __alloc_bootmem_node(pgdat_list
[bestnode
], pernodesize
,
391 PERCPU_PAGE_SIZE
, __pa(MAX_DMA_ADDRESS
));
397 * memory_less_nodes - allocate and initialize CPU only nodes pernode
400 static void __init
memory_less_nodes(void)
402 unsigned long pernodesize
;
406 for_each_node_mask(node
, memory_less_mask
) {
407 pernodesize
= compute_pernodesize(node
);
408 pernode
= memory_less_node_alloc(node
, pernodesize
);
409 fill_pernode(node
, __pa(pernode
), pernodesize
);
415 #ifdef CONFIG_SPARSEMEM
417 * register_sparse_mem - notify SPARSEMEM that this memory range exists.
418 * @start: physical start of range
419 * @end: physical end of range
422 * Simply calls SPARSEMEM to register memory section(s).
424 static int __init
register_sparse_mem(unsigned long start
, unsigned long end
,
429 start
= __pa(start
) >> PAGE_SHIFT
;
430 end
= __pa(end
) >> PAGE_SHIFT
;
431 nid
= early_pfn_to_nid(start
);
432 memory_present(nid
, start
, end
);
437 static void __init
arch_sparse_init(void)
439 efi_memmap_walk(register_sparse_mem
, NULL
);
443 #define arch_sparse_init() do {} while (0)
447 * find_memory - walk the EFI memory map and setup the bootmem allocator
449 * Called early in boot to setup the bootmem allocator, and to
450 * allocate the per-cpu and per-node structures.
452 void __init
find_memory(void)
458 if (num_online_nodes() == 0) {
459 printk(KERN_ERR
"node info missing!\n");
463 nodes_or(memory_less_mask
, memory_less_mask
, node_online_map
);
467 /* These actually end up getting called by call_pernode_memory() */
468 efi_memmap_walk(filter_rsvd_memory
, build_node_maps
);
469 efi_memmap_walk(filter_rsvd_memory
, find_pernode_space
);
471 for_each_online_node(node
)
472 if (mem_data
[node
].bootmem_data
.node_low_pfn
) {
473 node_clear(node
, memory_less_mask
);
474 mem_data
[node
].min_pfn
= ~0UL;
477 * Initialize the boot memory maps in reverse order since that's
478 * what the bootmem allocator expects
480 for (node
= MAX_NUMNODES
- 1; node
>= 0; node
--) {
481 unsigned long pernode
, pernodesize
, map
;
482 struct bootmem_data
*bdp
;
484 if (!node_online(node
))
486 else if (node_isset(node
, memory_less_mask
))
489 bdp
= &mem_data
[node
].bootmem_data
;
490 pernode
= mem_data
[node
].pernode_addr
;
491 pernodesize
= mem_data
[node
].pernode_size
;
492 map
= pernode
+ pernodesize
;
494 init_bootmem_node(pgdat_list
[node
],
496 bdp
->node_boot_start
>>PAGE_SHIFT
,
500 efi_memmap_walk(filter_rsvd_memory
, free_node_bootmem
);
502 reserve_pernode_space();
504 initialize_pernode_data();
506 max_pfn
= max_low_pfn
;
513 * per_cpu_init - setup per-cpu variables
515 * find_pernode_space() does most of this already, we just need to set
516 * local_per_cpu_offset
518 void __cpuinit
*per_cpu_init(void)
521 static int first_time
= 1;
524 if (smp_processor_id() != 0)
525 return __per_cpu_start
+ __per_cpu_offset
[smp_processor_id()];
529 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
530 per_cpu(local_per_cpu_offset
, cpu
) = __per_cpu_offset
[cpu
];
533 return __per_cpu_start
+ __per_cpu_offset
[smp_processor_id()];
535 #endif /* CONFIG_SMP */
538 * show_mem - give short summary of memory stats
540 * Shows a simple page count of reserved and used pages in the system.
541 * For discontig machines, it does this on a per-pgdat basis.
545 int i
, total_reserved
= 0;
546 int total_shared
= 0, total_cached
= 0;
547 unsigned long total_present
= 0;
550 printk("Mem-info:\n");
552 printk("Free swap: %6ldkB\n", nr_swap_pages
<<(PAGE_SHIFT
-10));
553 for_each_online_pgdat(pgdat
) {
554 unsigned long present
;
556 int shared
= 0, cached
= 0, reserved
= 0;
558 printk("Node ID: %d\n", pgdat
->node_id
);
559 pgdat_resize_lock(pgdat
, &flags
);
560 present
= pgdat
->node_present_pages
;
561 for(i
= 0; i
< pgdat
->node_spanned_pages
; i
++) {
563 if (pfn_valid(pgdat
->node_start_pfn
+ i
))
564 page
= pfn_to_page(pgdat
->node_start_pfn
+ i
);
566 i
= vmemmap_find_next_valid_pfn(pgdat
->node_id
,
570 if (PageReserved(page
))
572 else if (PageSwapCache(page
))
574 else if (page_count(page
))
575 shared
+= page_count(page
)-1;
577 pgdat_resize_unlock(pgdat
, &flags
);
578 total_present
+= present
;
579 total_reserved
+= reserved
;
580 total_cached
+= cached
;
581 total_shared
+= shared
;
582 printk("\t%ld pages of RAM\n", present
);
583 printk("\t%d reserved pages\n", reserved
);
584 printk("\t%d pages shared\n", shared
);
585 printk("\t%d pages swap cached\n", cached
);
587 printk("%ld pages of RAM\n", total_present
);
588 printk("%d reserved pages\n", total_reserved
);
589 printk("%d pages shared\n", total_shared
);
590 printk("%d pages swap cached\n", total_cached
);
591 printk("Total of %ld pages in page table cache\n",
592 pgtable_quicklist_total_size());
593 printk("%d free buffer pages\n", nr_free_buffer_pages());
597 * call_pernode_memory - use SRAT to call callback functions with node info
598 * @start: physical start of range
599 * @len: length of range
600 * @arg: function to call for each range
602 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
603 * out to which node a block of memory belongs. Ignore memory that we cannot
604 * identify, and split blocks that run across multiple nodes.
606 * Take this opportunity to round the start address up and the end address
607 * down to page boundaries.
609 void call_pernode_memory(unsigned long start
, unsigned long len
, void *arg
)
611 unsigned long rs
, re
, end
= start
+ len
;
612 void (*func
)(unsigned long, unsigned long, int);
615 start
= PAGE_ALIGN(start
);
622 if (!num_node_memblks
) {
623 /* No SRAT table, so assume one node (node 0) */
625 (*func
)(start
, end
- start
, 0);
629 for (i
= 0; i
< num_node_memblks
; i
++) {
630 rs
= max(start
, node_memblk
[i
].start_paddr
);
631 re
= min(end
, node_memblk
[i
].start_paddr
+
632 node_memblk
[i
].size
);
635 (*func
)(rs
, re
- rs
, node_memblk
[i
].nid
);
643 * count_node_pages - callback to build per-node memory info structures
644 * @start: physical start of range
645 * @len: length of range
646 * @node: node where this range resides
648 * Each node has it's own number of physical pages, DMAable pages, start, and
649 * end page frame number. This routine will be called by call_pernode_memory()
650 * for each piece of usable memory and will setup these values for each node.
651 * Very similar to build_maps().
653 static __init
int count_node_pages(unsigned long start
, unsigned long len
, int node
)
655 unsigned long end
= start
+ len
;
657 mem_data
[node
].num_physpages
+= len
>> PAGE_SHIFT
;
658 if (start
<= __pa(MAX_DMA_ADDRESS
))
659 mem_data
[node
].num_dma_physpages
+=
660 (min(end
, __pa(MAX_DMA_ADDRESS
)) - start
) >>PAGE_SHIFT
;
661 start
= GRANULEROUNDDOWN(start
);
662 start
= ORDERROUNDDOWN(start
);
663 end
= GRANULEROUNDUP(end
);
664 mem_data
[node
].max_pfn
= max(mem_data
[node
].max_pfn
,
666 mem_data
[node
].min_pfn
= min(mem_data
[node
].min_pfn
,
667 start
>> PAGE_SHIFT
);
673 * paging_init - setup page tables
675 * paging_init() sets up the page tables for each node of the system and frees
676 * the bootmem allocator memory for general use.
678 void __init
paging_init(void)
680 unsigned long max_dma
;
681 unsigned long zones_size
[MAX_NR_ZONES
];
682 unsigned long zholes_size
[MAX_NR_ZONES
];
683 unsigned long pfn_offset
= 0;
686 max_dma
= virt_to_phys((void *) MAX_DMA_ADDRESS
) >> PAGE_SHIFT
;
690 efi_memmap_walk(filter_rsvd_memory
, count_node_pages
);
692 #ifdef CONFIG_VIRTUAL_MEM_MAP
693 vmalloc_end
-= PAGE_ALIGN(ALIGN(max_low_pfn
, MAX_ORDER_NR_PAGES
) *
694 sizeof(struct page
));
695 vmem_map
= (struct page
*) vmalloc_end
;
696 efi_memmap_walk(create_mem_map_page_table
, NULL
);
697 printk("Virtual mem_map starts at 0x%p\n", vmem_map
);
700 for_each_online_node(node
) {
701 memset(zones_size
, 0, sizeof(zones_size
));
702 memset(zholes_size
, 0, sizeof(zholes_size
));
704 num_physpages
+= mem_data
[node
].num_physpages
;
706 if (mem_data
[node
].min_pfn
>= max_dma
) {
707 /* All of this node's memory is above ZONE_DMA */
708 zones_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
709 mem_data
[node
].min_pfn
;
710 zholes_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
711 mem_data
[node
].min_pfn
-
712 mem_data
[node
].num_physpages
;
713 } else if (mem_data
[node
].max_pfn
< max_dma
) {
714 /* All of this node's memory is in ZONE_DMA */
715 zones_size
[ZONE_DMA
] = mem_data
[node
].max_pfn
-
716 mem_data
[node
].min_pfn
;
717 zholes_size
[ZONE_DMA
] = mem_data
[node
].max_pfn
-
718 mem_data
[node
].min_pfn
-
719 mem_data
[node
].num_dma_physpages
;
721 /* This node has memory in both zones */
722 zones_size
[ZONE_DMA
] = max_dma
-
723 mem_data
[node
].min_pfn
;
724 zholes_size
[ZONE_DMA
] = zones_size
[ZONE_DMA
] -
725 mem_data
[node
].num_dma_physpages
;
726 zones_size
[ZONE_NORMAL
] = mem_data
[node
].max_pfn
-
728 zholes_size
[ZONE_NORMAL
] = zones_size
[ZONE_NORMAL
] -
729 (mem_data
[node
].num_physpages
-
730 mem_data
[node
].num_dma_physpages
);
733 pfn_offset
= mem_data
[node
].min_pfn
;
735 #ifdef CONFIG_VIRTUAL_MEM_MAP
736 NODE_DATA(node
)->node_mem_map
= vmem_map
+ pfn_offset
;
738 free_area_init_node(node
, NODE_DATA(node
), zones_size
,
739 pfn_offset
, zholes_size
);
742 zero_page_memmap_ptr
= virt_to_page(ia64_imva(empty_zero_page
));
745 pg_data_t
*arch_alloc_nodedata(int nid
)
747 unsigned long size
= compute_pernodesize(nid
);
749 return kzalloc(size
, GFP_KERNEL
);
752 void arch_free_nodedata(pg_data_t
*pgdat
)
757 void arch_refresh_nodedata(int update_node
, pg_data_t
*update_pgdat
)
759 pgdat_list
[update_node
] = update_pgdat
;