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 * memory_less_nodes - allocate and initialize CPU only nodes pernode
385 static void __init memory_less_nodes(void)
387 unsigned long pernodesize;
391 for_each_node_mask(node, memory_less_mask) {
392 pernodesize = compute_pernodesize(node);
393 pernode = memory_less_node_alloc(node, pernodesize);
394 fill_pernode(node, __pa(pernode), pernodesize);
400 #ifdef CONFIG_SPARSEMEM
402 * register_sparse_mem - notify SPARSEMEM that this memory range exists.
403 * @start: physical start of range
404 * @end: physical end of range
407 * Simply calls SPARSEMEM to register memory section(s).
409 static int __init register_sparse_mem(unsigned long start, unsigned long end,
414 start = __pa(start) >> PAGE_SHIFT;
415 end = __pa(end) >> PAGE_SHIFT;
416 nid = early_pfn_to_nid(start);
417 memory_present(nid, start, end);
422 static void __init arch_sparse_init(void)
424 efi_memmap_walk(register_sparse_mem, NULL);
428 #define arch_sparse_init() do {} while (0)
432 * find_memory - walk the EFI memory map and setup the bootmem allocator
434 * Called early in boot to setup the bootmem allocator, and to
435 * allocate the per-cpu and per-node structures.
437 void __init find_memory(void)
443 if (num_online_nodes() == 0) {
444 printk(KERN_ERR "node info missing!\n");
448 nodes_or(memory_less_mask, memory_less_mask, node_online_map);
452 /* These actually end up getting called by call_pernode_memory() */
453 efi_memmap_walk(filter_rsvd_memory, build_node_maps);
454 efi_memmap_walk(filter_rsvd_memory, find_pernode_space);
456 for_each_online_node(node)
457 if (mem_data[node].bootmem_data.node_low_pfn) {
458 node_clear(node, memory_less_mask);
459 mem_data[node].min_pfn = ~0UL;
462 * Initialize the boot memory maps in reverse order since that's
463 * what the bootmem allocator expects
465 for (node = MAX_NUMNODES - 1; node >= 0; node--) {
466 unsigned long pernode, pernodesize, map;
467 struct bootmem_data *bdp;
469 if (!node_online(node))
471 else if (node_isset(node, memory_less_mask))
474 bdp = &mem_data[node].bootmem_data;
475 pernode = mem_data[node].pernode_addr;
476 pernodesize = mem_data[node].pernode_size;
477 map = pernode + pernodesize;
479 init_bootmem_node(mem_data[node].pgdat,
481 bdp->node_boot_start>>PAGE_SHIFT,
485 efi_memmap_walk(filter_rsvd_memory, free_node_bootmem);
487 reserve_pernode_space();
489 initialize_pernode_data();
491 max_pfn = max_low_pfn;
498 * per_cpu_init - setup per-cpu variables
500 * find_pernode_space() does most of this already, we just need to set
501 * local_per_cpu_offset
503 void __cpuinit *per_cpu_init(void)
506 static int first_time = 1;
509 if (smp_processor_id() != 0)
510 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
514 for (cpu = 0; cpu < NR_CPUS; cpu++)
515 per_cpu(local_per_cpu_offset, cpu) = __per_cpu_offset[cpu];
518 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
520 #endif /* CONFIG_SMP */
522 #ifdef CONFIG_VIRTUAL_MEM_MAP
523 static inline int find_next_valid_pfn_for_pgdat(pg_data_t *pgdat, int i)
525 unsigned long end_address, hole_next_pfn;
526 unsigned long stop_address;
528 end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
529 end_address = PAGE_ALIGN(end_address);
531 stop_address = (unsigned long) &vmem_map[
532 pgdat->node_start_pfn + pgdat->node_spanned_pages];
540 pgd = pgd_offset_k(end_address);
541 if (pgd_none(*pgd)) {
542 end_address += PGDIR_SIZE;
546 pud = pud_offset(pgd, end_address);
547 if (pud_none(*pud)) {
548 end_address += PUD_SIZE;
552 pmd = pmd_offset(pud, end_address);
553 if (pmd_none(*pmd)) {
554 end_address += PMD_SIZE;
558 pte = pte_offset_kernel(pmd, end_address);
560 if (pte_none(*pte)) {
561 end_address += PAGE_SIZE;
563 if ((end_address < stop_address) &&
564 (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
568 /* Found next valid vmem_map page */
570 } while (end_address < stop_address);
572 end_address = min(end_address, stop_address);
573 end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
574 hole_next_pfn = end_address / sizeof(struct page);
575 return hole_next_pfn - pgdat->node_start_pfn;
578 static inline int find_next_valid_pfn_for_pgdat(pg_data_t *pgdat, int i)
585 * show_mem - give short summary of memory stats
587 * Shows a simple page count of reserved and used pages in the system.
588 * For discontig machines, it does this on a per-pgdat basis.
592 int i, total_reserved = 0;
593 int total_shared = 0, total_cached = 0;
594 unsigned long total_present = 0;
597 printk("Mem-info:\n");
599 printk("Free swap: %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
600 for_each_online_pgdat(pgdat) {
601 unsigned long present;
603 int shared = 0, cached = 0, reserved = 0;
605 printk("Node ID: %d\n", pgdat->node_id);
606 pgdat_resize_lock(pgdat, &flags);
607 present = pgdat->node_present_pages;
608 for(i = 0; i < pgdat->node_spanned_pages; i++) {
610 if (pfn_valid(pgdat->node_start_pfn + i))
611 page = pfn_to_page(pgdat->node_start_pfn + i);
613 i = find_next_valid_pfn_for_pgdat(pgdat, i) - 1;
616 if (PageReserved(page))
618 else if (PageSwapCache(page))
620 else if (page_count(page))
621 shared += page_count(page)-1;
623 pgdat_resize_unlock(pgdat, &flags);
624 total_present += present;
625 total_reserved += reserved;
626 total_cached += cached;
627 total_shared += shared;
628 printk("\t%ld pages of RAM\n", present);
629 printk("\t%d reserved pages\n", reserved);
630 printk("\t%d pages shared\n", shared);
631 printk("\t%d pages swap cached\n", cached);
633 printk("%ld pages of RAM\n", total_present);
634 printk("%d reserved pages\n", total_reserved);
635 printk("%d pages shared\n", total_shared);
636 printk("%d pages swap cached\n", total_cached);
637 printk("Total of %ld pages in page table cache\n",
638 pgtable_quicklist_total_size());
639 printk("%d free buffer pages\n", nr_free_buffer_pages());
643 * call_pernode_memory - use SRAT to call callback functions with node info
644 * @start: physical start of range
645 * @len: length of range
646 * @arg: function to call for each range
648 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
649 * out to which node a block of memory belongs. Ignore memory that we cannot
650 * identify, and split blocks that run across multiple nodes.
652 * Take this opportunity to round the start address up and the end address
653 * down to page boundaries.
655 void call_pernode_memory(unsigned long start, unsigned long len, void *arg)
657 unsigned long rs, re, end = start + len;
658 void (*func)(unsigned long, unsigned long, int);
661 start = PAGE_ALIGN(start);
668 if (!num_node_memblks) {
669 /* No SRAT table, so assume one node (node 0) */
671 (*func)(start, end - start, 0);
675 for (i = 0; i < num_node_memblks; i++) {
676 rs = max(start, node_memblk[i].start_paddr);
677 re = min(end, node_memblk[i].start_paddr +
678 node_memblk[i].size);
681 (*func)(rs, re - rs, node_memblk[i].nid);
689 * count_node_pages - callback to build per-node memory info structures
690 * @start: physical start of range
691 * @len: length of range
692 * @node: node where this range resides
694 * Each node has it's own number of physical pages, DMAable pages, start, and
695 * end page frame number. This routine will be called by call_pernode_memory()
696 * for each piece of usable memory and will setup these values for each node.
697 * Very similar to build_maps().
699 static __init int count_node_pages(unsigned long start, unsigned long len, int node)
701 unsigned long end = start + len;
703 mem_data[node].num_physpages += len >> PAGE_SHIFT;
704 if (start <= __pa(MAX_DMA_ADDRESS))
705 mem_data[node].num_dma_physpages +=
706 (min(end, __pa(MAX_DMA_ADDRESS)) - start) >>PAGE_SHIFT;
707 start = GRANULEROUNDDOWN(start);
708 start = ORDERROUNDDOWN(start);
709 end = GRANULEROUNDUP(end);
710 mem_data[node].max_pfn = max(mem_data[node].max_pfn,
712 mem_data[node].min_pfn = min(mem_data[node].min_pfn,
713 start >> PAGE_SHIFT);
719 * paging_init - setup page tables
721 * paging_init() sets up the page tables for each node of the system and frees
722 * the bootmem allocator memory for general use.
724 void __init paging_init(void)
726 unsigned long max_dma;
727 unsigned long zones_size[MAX_NR_ZONES];
728 unsigned long zholes_size[MAX_NR_ZONES];
729 unsigned long pfn_offset = 0;
732 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
736 efi_memmap_walk(filter_rsvd_memory, count_node_pages);
738 #ifdef CONFIG_VIRTUAL_MEM_MAP
739 vmalloc_end -= PAGE_ALIGN(max_low_pfn * sizeof(struct page));
740 vmem_map = (struct page *) vmalloc_end;
741 efi_memmap_walk(create_mem_map_page_table, NULL);
742 printk("Virtual mem_map starts at 0x%p\n", vmem_map);
745 for_each_online_node(node) {
746 memset(zones_size, 0, sizeof(zones_size));
747 memset(zholes_size, 0, sizeof(zholes_size));
749 num_physpages += mem_data[node].num_physpages;
751 if (mem_data[node].min_pfn >= max_dma) {
752 /* All of this node's memory is above ZONE_DMA */
753 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
754 mem_data[node].min_pfn;
755 zholes_size[ZONE_NORMAL] = mem_data[node].max_pfn -
756 mem_data[node].min_pfn -
757 mem_data[node].num_physpages;
758 } else if (mem_data[node].max_pfn < max_dma) {
759 /* All of this node's memory is in ZONE_DMA */
760 zones_size[ZONE_DMA] = mem_data[node].max_pfn -
761 mem_data[node].min_pfn;
762 zholes_size[ZONE_DMA] = mem_data[node].max_pfn -
763 mem_data[node].min_pfn -
764 mem_data[node].num_dma_physpages;
766 /* This node has memory in both zones */
767 zones_size[ZONE_DMA] = max_dma -
768 mem_data[node].min_pfn;
769 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
770 mem_data[node].num_dma_physpages;
771 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
773 zholes_size[ZONE_NORMAL] = zones_size[ZONE_NORMAL] -
774 (mem_data[node].num_physpages -
775 mem_data[node].num_dma_physpages);
778 pfn_offset = mem_data[node].min_pfn;
780 #ifdef CONFIG_VIRTUAL_MEM_MAP
781 NODE_DATA(node)->node_mem_map = vmem_map + pfn_offset;
783 free_area_init_node(node, NODE_DATA(node), zones_size,
784 pfn_offset, zholes_size);
787 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page));