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)
316 for_each_online_node(node) {
317 dst = LOCAL_DATA_ADDR(pgdat_list[node])->pg_data_ptrs;
318 memcpy(dst, pgdat_list, sizeof(pgdat_list));
323 * initialize_pernode_data - fixup per-cpu & per-node pointers
325 * Each node's per-node area has a copy of the global pg_data_t list, so
326 * we copy that to each node here, as well as setting the per-cpu pointer
327 * to the local node data structure. The active_cpus field of the per-node
328 * structure gets setup by the platform_cpu_init() function later.
330 static void __init initialize_pernode_data(void)
337 /* Set the node_data pointer for each per-cpu struct */
338 for (cpu = 0; cpu < NR_CPUS; cpu++) {
339 node = node_cpuid[cpu].nid;
340 per_cpu(cpu_info, cpu).node_data = mem_data[node].node_data;
344 struct cpuinfo_ia64 *cpu0_cpu_info;
346 node = node_cpuid[cpu].nid;
347 cpu0_cpu_info = (struct cpuinfo_ia64 *)(__phys_per_cpu_start +
348 ((char *)&per_cpu__cpu_info - __per_cpu_start));
349 cpu0_cpu_info->node_data = mem_data[node].node_data;
351 #endif /* CONFIG_SMP */
355 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit
356 * node but fall back to any other node when __alloc_bootmem_node fails
359 * @pernodesize: size of this node's pernode data
361 static void __init *memory_less_node_alloc(int nid, unsigned long pernodesize)
365 int bestnode = -1, node, anynode = 0;
367 for_each_online_node(node) {
368 if (node_isset(node, memory_less_mask))
370 else if (node_distance(nid, node) < best) {
371 best = node_distance(nid, node);
380 ptr = __alloc_bootmem_node(pgdat_list[bestnode], pernodesize,
381 PERCPU_PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
387 * memory_less_nodes - allocate and initialize CPU only nodes pernode
390 static void __init memory_less_nodes(void)
392 unsigned long pernodesize;
396 for_each_node_mask(node, memory_less_mask) {
397 pernodesize = compute_pernodesize(node);
398 pernode = memory_less_node_alloc(node, pernodesize);
399 fill_pernode(node, __pa(pernode), pernodesize);
405 #ifdef CONFIG_SPARSEMEM
407 * register_sparse_mem - notify SPARSEMEM that this memory range exists.
408 * @start: physical start of range
409 * @end: physical end of range
412 * Simply calls SPARSEMEM to register memory section(s).
414 static int __init register_sparse_mem(unsigned long start, unsigned long end,
419 start = __pa(start) >> PAGE_SHIFT;
420 end = __pa(end) >> PAGE_SHIFT;
421 nid = early_pfn_to_nid(start);
422 memory_present(nid, start, end);
427 static void __init arch_sparse_init(void)
429 efi_memmap_walk(register_sparse_mem, NULL);
433 #define arch_sparse_init() do {} while (0)
437 * find_memory - walk the EFI memory map and setup the bootmem allocator
439 * Called early in boot to setup the bootmem allocator, and to
440 * allocate the per-cpu and per-node structures.
442 void __init find_memory(void)
448 if (num_online_nodes() == 0) {
449 printk(KERN_ERR "node info missing!\n");
453 nodes_or(memory_less_mask, memory_less_mask, node_online_map);
457 /* These actually end up getting called by call_pernode_memory() */
458 efi_memmap_walk(filter_rsvd_memory, build_node_maps);
459 efi_memmap_walk(filter_rsvd_memory, find_pernode_space);
461 for_each_online_node(node)
462 if (mem_data[node].bootmem_data.node_low_pfn) {
463 node_clear(node, memory_less_mask);
464 mem_data[node].min_pfn = ~0UL;
467 * Initialize the boot memory maps in reverse order since that's
468 * what the bootmem allocator expects
470 for (node = MAX_NUMNODES - 1; node >= 0; node--) {
471 unsigned long pernode, pernodesize, map;
472 struct bootmem_data *bdp;
474 if (!node_online(node))
476 else if (node_isset(node, memory_less_mask))
479 bdp = &mem_data[node].bootmem_data;
480 pernode = mem_data[node].pernode_addr;
481 pernodesize = mem_data[node].pernode_size;
482 map = pernode + pernodesize;
484 init_bootmem_node(pgdat_list[node],
486 bdp->node_boot_start>>PAGE_SHIFT,
490 efi_memmap_walk(filter_rsvd_memory, free_node_bootmem);
492 reserve_pernode_space();
494 initialize_pernode_data();
496 max_pfn = max_low_pfn;
503 * per_cpu_init - setup per-cpu variables
505 * find_pernode_space() does most of this already, we just need to set
506 * local_per_cpu_offset
508 void __cpuinit *per_cpu_init(void)
511 static int first_time = 1;
514 if (smp_processor_id() != 0)
515 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
519 for (cpu = 0; cpu < NR_CPUS; cpu++)
520 per_cpu(local_per_cpu_offset, cpu) = __per_cpu_offset[cpu];
523 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
525 #endif /* CONFIG_SMP */
527 #ifdef CONFIG_VIRTUAL_MEM_MAP
528 static inline int find_next_valid_pfn_for_pgdat(pg_data_t *pgdat, int i)
530 unsigned long end_address, hole_next_pfn;
531 unsigned long stop_address;
533 end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
534 end_address = PAGE_ALIGN(end_address);
536 stop_address = (unsigned long) &vmem_map[
537 pgdat->node_start_pfn + pgdat->node_spanned_pages];
545 pgd = pgd_offset_k(end_address);
546 if (pgd_none(*pgd)) {
547 end_address += PGDIR_SIZE;
551 pud = pud_offset(pgd, end_address);
552 if (pud_none(*pud)) {
553 end_address += PUD_SIZE;
557 pmd = pmd_offset(pud, end_address);
558 if (pmd_none(*pmd)) {
559 end_address += PMD_SIZE;
563 pte = pte_offset_kernel(pmd, end_address);
565 if (pte_none(*pte)) {
566 end_address += PAGE_SIZE;
568 if ((end_address < stop_address) &&
569 (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
573 /* Found next valid vmem_map page */
575 } while (end_address < stop_address);
577 end_address = min(end_address, stop_address);
578 end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
579 hole_next_pfn = end_address / sizeof(struct page);
580 return hole_next_pfn - pgdat->node_start_pfn;
583 static inline int find_next_valid_pfn_for_pgdat(pg_data_t *pgdat, int i)
590 * show_mem - give short summary of memory stats
592 * Shows a simple page count of reserved and used pages in the system.
593 * For discontig machines, it does this on a per-pgdat basis.
597 int i, total_reserved = 0;
598 int total_shared = 0, total_cached = 0;
599 unsigned long total_present = 0;
602 printk("Mem-info:\n");
604 printk("Free swap: %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
605 for_each_online_pgdat(pgdat) {
606 unsigned long present;
608 int shared = 0, cached = 0, reserved = 0;
610 printk("Node ID: %d\n", pgdat->node_id);
611 pgdat_resize_lock(pgdat, &flags);
612 present = pgdat->node_present_pages;
613 for(i = 0; i < pgdat->node_spanned_pages; i++) {
615 if (pfn_valid(pgdat->node_start_pfn + i))
616 page = pfn_to_page(pgdat->node_start_pfn + i);
618 i = find_next_valid_pfn_for_pgdat(pgdat, i) - 1;
621 if (PageReserved(page))
623 else if (PageSwapCache(page))
625 else if (page_count(page))
626 shared += page_count(page)-1;
628 pgdat_resize_unlock(pgdat, &flags);
629 total_present += present;
630 total_reserved += reserved;
631 total_cached += cached;
632 total_shared += shared;
633 printk("\t%ld pages of RAM\n", present);
634 printk("\t%d reserved pages\n", reserved);
635 printk("\t%d pages shared\n", shared);
636 printk("\t%d pages swap cached\n", cached);
638 printk("%ld pages of RAM\n", total_present);
639 printk("%d reserved pages\n", total_reserved);
640 printk("%d pages shared\n", total_shared);
641 printk("%d pages swap cached\n", total_cached);
642 printk("Total of %ld pages in page table cache\n",
643 pgtable_quicklist_total_size());
644 printk("%d free buffer pages\n", nr_free_buffer_pages());
648 * call_pernode_memory - use SRAT to call callback functions with node info
649 * @start: physical start of range
650 * @len: length of range
651 * @arg: function to call for each range
653 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
654 * out to which node a block of memory belongs. Ignore memory that we cannot
655 * identify, and split blocks that run across multiple nodes.
657 * Take this opportunity to round the start address up and the end address
658 * down to page boundaries.
660 void call_pernode_memory(unsigned long start, unsigned long len, void *arg)
662 unsigned long rs, re, end = start + len;
663 void (*func)(unsigned long, unsigned long, int);
666 start = PAGE_ALIGN(start);
673 if (!num_node_memblks) {
674 /* No SRAT table, so assume one node (node 0) */
676 (*func)(start, end - start, 0);
680 for (i = 0; i < num_node_memblks; i++) {
681 rs = max(start, node_memblk[i].start_paddr);
682 re = min(end, node_memblk[i].start_paddr +
683 node_memblk[i].size);
686 (*func)(rs, re - rs, node_memblk[i].nid);
694 * count_node_pages - callback to build per-node memory info structures
695 * @start: physical start of range
696 * @len: length of range
697 * @node: node where this range resides
699 * Each node has it's own number of physical pages, DMAable pages, start, and
700 * end page frame number. This routine will be called by call_pernode_memory()
701 * for each piece of usable memory and will setup these values for each node.
702 * Very similar to build_maps().
704 static __init int count_node_pages(unsigned long start, unsigned long len, int node)
706 unsigned long end = start + len;
708 mem_data[node].num_physpages += len >> PAGE_SHIFT;
709 if (start <= __pa(MAX_DMA_ADDRESS))
710 mem_data[node].num_dma_physpages +=
711 (min(end, __pa(MAX_DMA_ADDRESS)) - start) >>PAGE_SHIFT;
712 start = GRANULEROUNDDOWN(start);
713 start = ORDERROUNDDOWN(start);
714 end = GRANULEROUNDUP(end);
715 mem_data[node].max_pfn = max(mem_data[node].max_pfn,
717 mem_data[node].min_pfn = min(mem_data[node].min_pfn,
718 start >> PAGE_SHIFT);
724 * paging_init - setup page tables
726 * paging_init() sets up the page tables for each node of the system and frees
727 * the bootmem allocator memory for general use.
729 void __init paging_init(void)
731 unsigned long max_dma;
732 unsigned long zones_size[MAX_NR_ZONES];
733 unsigned long zholes_size[MAX_NR_ZONES];
734 unsigned long pfn_offset = 0;
737 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
741 efi_memmap_walk(filter_rsvd_memory, count_node_pages);
743 #ifdef CONFIG_VIRTUAL_MEM_MAP
744 vmalloc_end -= PAGE_ALIGN(max_low_pfn * sizeof(struct page));
745 vmem_map = (struct page *) vmalloc_end;
746 efi_memmap_walk(create_mem_map_page_table, NULL);
747 printk("Virtual mem_map starts at 0x%p\n", vmem_map);
750 for_each_online_node(node) {
751 memset(zones_size, 0, sizeof(zones_size));
752 memset(zholes_size, 0, sizeof(zholes_size));
754 num_physpages += mem_data[node].num_physpages;
756 if (mem_data[node].min_pfn >= max_dma) {
757 /* All of this node's memory is above ZONE_DMA */
758 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
759 mem_data[node].min_pfn;
760 zholes_size[ZONE_NORMAL] = mem_data[node].max_pfn -
761 mem_data[node].min_pfn -
762 mem_data[node].num_physpages;
763 } else if (mem_data[node].max_pfn < max_dma) {
764 /* All of this node's memory is in ZONE_DMA */
765 zones_size[ZONE_DMA] = mem_data[node].max_pfn -
766 mem_data[node].min_pfn;
767 zholes_size[ZONE_DMA] = mem_data[node].max_pfn -
768 mem_data[node].min_pfn -
769 mem_data[node].num_dma_physpages;
771 /* This node has memory in both zones */
772 zones_size[ZONE_DMA] = max_dma -
773 mem_data[node].min_pfn;
774 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
775 mem_data[node].num_dma_physpages;
776 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
778 zholes_size[ZONE_NORMAL] = zones_size[ZONE_NORMAL] -
779 (mem_data[node].num_physpages -
780 mem_data[node].num_dma_physpages);
783 pfn_offset = mem_data[node].min_pfn;
785 #ifdef CONFIG_VIRTUAL_MEM_MAP
786 NODE_DATA(node)->node_mem_map = vmem_map + pfn_offset;
788 free_area_init_node(node, NODE_DATA(node), zones_size,
789 pfn_offset, zholes_size);
792 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page));
795 pg_data_t *arch_alloc_nodedata(int nid)
797 unsigned long size = compute_pernodesize(nid);
799 return kzalloc(size, GFP_KERNEL);
802 void arch_free_nodedata(pg_data_t *pgdat)
807 void arch_refresh_nodedata(int update_node, pg_data_t *update_pgdat)
809 pgdat_list[update_node] = update_pgdat;