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>
10 * Platform initialization for Discontig Memory
13 #include <linux/kernel.h>
15 #include <linux/swap.h>
16 #include <linux/bootmem.h>
17 #include <linux/acpi.h>
18 #include <linux/efi.h>
19 #include <asm/pgalloc.h>
21 #include <asm/meminit.h>
23 #include <asm/sections.h>
26 * Track per-node information needed to setup the boot memory allocator, the
27 * per-node areas, and the real VM.
29 struct early_node_data {
30 struct ia64_node_data *node_data;
32 unsigned long pernode_addr;
33 unsigned long pernode_size;
34 struct bootmem_data bootmem_data;
35 unsigned long num_physpages;
36 unsigned long num_dma_physpages;
37 unsigned long min_pfn;
38 unsigned long max_pfn;
41 static struct early_node_data mem_data[NR_NODES] __initdata;
44 * reassign_cpu_only_nodes - called from find_memory to move CPU-only nodes to a memory node
46 * This function will move nodes with only CPUs (no memory)
47 * to a node with memory which is at the minimum numa_slit distance.
48 * Any reassigments will result in the compression of the nodes
49 * and renumbering the nid values where appropriate.
50 * The static declarations below are to avoid large stack size which
51 * makes the code not re-entrant.
53 static void __init reassign_cpu_only_nodes(void)
55 struct node_memblk_s *p;
56 int i, j, k, nnode, nid, cpu, cpunid;
58 static DECLARE_BITMAP(nodes_with_mem, NR_NODES) __initdata;
59 static u8 numa_slit_fix[MAX_NUMNODES * MAX_NUMNODES] __initdata;
60 static int node_flip[NR_NODES] __initdata;
62 for (nnode = 0, p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
63 if (!test_bit(p->nid, (void *) nodes_with_mem)) {
64 set_bit(p->nid, (void *) nodes_with_mem);
69 * All nids with memory.
71 if (nnode == numnodes)
75 * Change nids and attempt to migrate CPU-only nodes
76 * to the best numa_slit (closest neighbor) possible.
77 * For reassigned CPU nodes a nid can't be arrived at
78 * until after this loop because the target nid's new
79 * identity might not have been established yet. So
80 * new nid values are fabricated above numnodes and
81 * mapped back later to their true value.
83 for (nid = 0, i = 0; i < numnodes; i++) {
84 if (test_bit(i, (void *) nodes_with_mem)) {
86 * Save original nid value for numa_slit
87 * fixup and node_cpuid reassignments.
96 for (p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
105 for (cpu = 0; cpu < NR_CPUS; cpu++)
106 if (node_cpuid[cpu].nid == i) {
107 /* For nodes not being reassigned just fix the cpu's nid. */
108 if (cpunid < numnodes) {
109 node_cpuid[cpu].nid = cpunid;
114 * For nodes being reassigned, find best node by
115 * numa_slit information and then make a temporary
116 * nid value based on current nid and numnodes.
118 for (slit = 0xff, k = numnodes + numnodes, j = 0; j < numnodes; j++)
121 else if (test_bit(j, (void *) nodes_with_mem)) {
122 cslit = numa_slit[i * numnodes + j];
129 node_cpuid[cpu].nid = k;
134 * Fixup temporary nid values for CPU-only nodes.
136 for (cpu = 0; cpu < NR_CPUS; cpu++)
137 if (node_cpuid[cpu].nid == (numnodes + numnodes))
138 node_cpuid[cpu].nid = nnode - 1;
140 for (i = 0; i < nnode; i++)
141 if (node_flip[i] == (node_cpuid[cpu].nid - numnodes)) {
142 node_cpuid[cpu].nid = i;
147 * Fix numa_slit by compressing from larger
148 * nid array to reduced nid array.
150 for (i = 0; i < nnode; i++)
151 for (j = 0; j < nnode; j++)
152 numa_slit_fix[i * nnode + j] =
153 numa_slit[node_flip[i] * numnodes + node_flip[j]];
155 memcpy(numa_slit, numa_slit_fix, sizeof (numa_slit));
157 for (i = nnode; i < numnodes; i++)
166 * To prevent cache aliasing effects, align per-node structures so that they
167 * start at addresses that are strided by node number.
169 #define NODEDATA_ALIGN(addr, node) \
170 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + (node)*PERCPU_PAGE_SIZE)
173 * build_node_maps - callback to setup bootmem structs for each node
174 * @start: physical start of range
175 * @len: length of range
176 * @node: node where this range resides
178 * We allocate a struct bootmem_data for each piece of memory that we wish to
179 * treat as a virtually contiguous block (i.e. each node). Each such block
180 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
181 * if necessary. Any non-existent pages will simply be part of the virtual
182 * memmap. We also update min_low_pfn and max_low_pfn here as we receive
183 * memory ranges from the caller.
185 static int __init build_node_maps(unsigned long start, unsigned long len,
188 unsigned long cstart, epfn, end = start + len;
189 struct bootmem_data *bdp = &mem_data[node].bootmem_data;
191 epfn = GRANULEROUNDUP(end) >> PAGE_SHIFT;
192 cstart = GRANULEROUNDDOWN(start);
194 if (!bdp->node_low_pfn) {
195 bdp->node_boot_start = cstart;
196 bdp->node_low_pfn = epfn;
198 bdp->node_boot_start = min(cstart, bdp->node_boot_start);
199 bdp->node_low_pfn = max(epfn, bdp->node_low_pfn);
202 min_low_pfn = min(min_low_pfn, bdp->node_boot_start>>PAGE_SHIFT);
203 max_low_pfn = max(max_low_pfn, bdp->node_low_pfn);
209 * early_nr_cpus_node - return number of cpus on a given node
210 * @node: node to check
212 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because
213 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
216 static int early_nr_cpus_node(int node)
220 for (cpu = 0; cpu < NR_CPUS; cpu++)
221 if (node == node_cpuid[cpu].nid)
228 * find_pernode_space - allocate memory for memory map and per-node structures
229 * @start: physical start of range
230 * @len: length of range
231 * @node: node where this range resides
233 * This routine reserves space for the per-cpu data struct, the list of
234 * pg_data_ts and the per-node data struct. Each node will have something like
235 * the following in the first chunk of addr. space large enough to hold it.
237 * ________________________
239 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
240 * | PERCPU_PAGE_SIZE * | start and length big enough
242 * |------------------------|
243 * | local pg_data_t * |
244 * |------------------------|
245 * | local ia64_node_data |
246 * |------------------------|
248 * |________________________|
250 * Once this space has been set aside, the bootmem maps are initialized. We
251 * could probably move the allocation of the per-cpu and ia64_node_data space
252 * outside of this function and use alloc_bootmem_node(), but doing it here
253 * is straightforward and we get the alignments we want so...
255 static int __init find_pernode_space(unsigned long start, unsigned long len,
258 unsigned long epfn, cpu, cpus;
259 unsigned long pernodesize = 0, pernode, pages, mapsize;
261 struct bootmem_data *bdp = &mem_data[node].bootmem_data;
263 epfn = (start + len) >> PAGE_SHIFT;
265 pages = bdp->node_low_pfn - (bdp->node_boot_start >> PAGE_SHIFT);
266 mapsize = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
269 * Make sure this memory falls within this node's usable memory
270 * since we may have thrown some away in build_maps().
272 if (start < bdp->node_boot_start || epfn > bdp->node_low_pfn)
275 /* Don't setup this node's local space twice... */
276 if (mem_data[node].pernode_addr)
280 * Calculate total size needed, incl. what's necessary
281 * for good alignment and alias prevention.
283 cpus = early_nr_cpus_node(node);
284 pernodesize += PERCPU_PAGE_SIZE * cpus;
285 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
286 pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
287 pernodesize = PAGE_ALIGN(pernodesize);
288 pernode = NODEDATA_ALIGN(start, node);
290 /* Is this range big enough for what we want to store here? */
291 if (start + len > (pernode + pernodesize + mapsize)) {
292 mem_data[node].pernode_addr = pernode;
293 mem_data[node].pernode_size = pernodesize;
294 memset(__va(pernode), 0, pernodesize);
296 cpu_data = (void *)pernode;
297 pernode += PERCPU_PAGE_SIZE * cpus;
299 mem_data[node].pgdat = __va(pernode);
300 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
302 mem_data[node].node_data = __va(pernode);
303 pernode += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
305 mem_data[node].pgdat->bdata = bdp;
306 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
309 * Copy the static per-cpu data into the region we
310 * just set aside and then setup __per_cpu_offset
311 * for each CPU on this node.
313 for (cpu = 0; cpu < NR_CPUS; cpu++) {
314 if (node == node_cpuid[cpu].nid) {
315 memcpy(__va(cpu_data), __phys_per_cpu_start,
316 __per_cpu_end - __per_cpu_start);
317 __per_cpu_offset[cpu] = (char*)__va(cpu_data) -
319 cpu_data += PERCPU_PAGE_SIZE;
328 * free_node_bootmem - free bootmem allocator memory for use
329 * @start: physical start of range
330 * @len: length of range
331 * @node: node where this range resides
333 * Simply calls the bootmem allocator to free the specified ranged from
334 * the given pg_data_t's bdata struct. After this function has been called
335 * for all the entries in the EFI memory map, the bootmem allocator will
336 * be ready to service allocation requests.
338 static int __init free_node_bootmem(unsigned long start, unsigned long len,
341 free_bootmem_node(mem_data[node].pgdat, start, len);
347 * reserve_pernode_space - reserve memory for per-node space
349 * Reserve the space used by the bootmem maps & per-node space in the boot
350 * allocator so that when we actually create the real mem maps we don't
353 static void __init reserve_pernode_space(void)
355 unsigned long base, size, pages;
356 struct bootmem_data *bdp;
359 for (node = 0; node < numnodes; node++) {
360 pg_data_t *pdp = mem_data[node].pgdat;
364 /* First the bootmem_map itself */
365 pages = bdp->node_low_pfn - (bdp->node_boot_start>>PAGE_SHIFT);
366 size = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
367 base = __pa(bdp->node_bootmem_map);
368 reserve_bootmem_node(pdp, base, size);
370 /* Now the per-node space */
371 size = mem_data[node].pernode_size;
372 base = __pa(mem_data[node].pernode_addr);
373 reserve_bootmem_node(pdp, base, size);
378 * initialize_pernode_data - fixup per-cpu & per-node pointers
380 * Each node's per-node area has a copy of the global pg_data_t list, so
381 * we copy that to each node here, as well as setting the per-cpu pointer
382 * to the local node data structure. The active_cpus field of the per-node
383 * structure gets setup by the platform_cpu_init() function later.
385 static void __init initialize_pernode_data(void)
388 pg_data_t *pgdat_list[NR_NODES];
390 for (node = 0; node < numnodes; node++)
391 pgdat_list[node] = mem_data[node].pgdat;
393 /* Copy the pg_data_t list to each node and init the node field */
394 for (node = 0; node < numnodes; node++) {
395 memcpy(mem_data[node].node_data->pg_data_ptrs, pgdat_list,
399 /* Set the node_data pointer for each per-cpu struct */
400 for (cpu = 0; cpu < NR_CPUS; cpu++) {
401 node = node_cpuid[cpu].nid;
402 per_cpu(cpu_info, cpu).node_data = mem_data[node].node_data;
407 * find_memory - walk the EFI memory map and setup the bootmem allocator
409 * Called early in boot to setup the bootmem allocator, and to
410 * allocate the per-cpu and per-node structures.
412 void __init find_memory(void)
419 printk(KERN_ERR "node info missing!\n");
427 reassign_cpu_only_nodes();
429 /* These actually end up getting called by call_pernode_memory() */
430 efi_memmap_walk(filter_rsvd_memory, build_node_maps);
431 efi_memmap_walk(filter_rsvd_memory, find_pernode_space);
434 * Initialize the boot memory maps in reverse order since that's
435 * what the bootmem allocator expects
437 for (node = numnodes - 1; node >= 0; node--) {
438 unsigned long pernode, pernodesize, map;
439 struct bootmem_data *bdp;
441 bdp = &mem_data[node].bootmem_data;
442 pernode = mem_data[node].pernode_addr;
443 pernodesize = mem_data[node].pernode_size;
444 map = pernode + pernodesize;
446 /* Sanity check... */
448 panic("pernode space for node %d "
449 "could not be allocated!", node);
451 init_bootmem_node(mem_data[node].pgdat,
453 bdp->node_boot_start>>PAGE_SHIFT,
457 efi_memmap_walk(filter_rsvd_memory, free_node_bootmem);
459 reserve_pernode_space();
460 initialize_pernode_data();
462 max_pfn = max_low_pfn;
468 * per_cpu_init - setup per-cpu variables
470 * find_pernode_space() does most of this already, we just need to set
471 * local_per_cpu_offset
473 void *per_cpu_init(void)
477 if (smp_processor_id() == 0) {
478 for (cpu = 0; cpu < NR_CPUS; cpu++) {
479 per_cpu(local_per_cpu_offset, cpu) =
480 __per_cpu_offset[cpu];
484 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
488 * show_mem - give short summary of memory stats
490 * Shows a simple page count of reserved and used pages in the system.
491 * For discontig machines, it does this on a per-pgdat basis.
496 int shared = 0, cached = 0;
499 printk("Mem-info:\n");
501 printk("Free swap: %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
502 for_each_pgdat(pgdat) {
503 printk("Node ID: %d\n", pgdat->node_id);
504 for(i = 0; i < pgdat->node_spanned_pages; i++) {
505 if (!ia64_pfn_valid(pgdat->node_start_pfn+i))
507 if (PageReserved(pgdat->node_mem_map+i))
509 else if (PageSwapCache(pgdat->node_mem_map+i))
511 else if (page_count(pgdat->node_mem_map+i))
512 shared += page_count(pgdat->node_mem_map+i)-1;
514 printk("\t%ld pages of RAM\n", pgdat->node_present_pages);
515 printk("\t%d reserved pages\n", reserved);
516 printk("\t%d pages shared\n", shared);
517 printk("\t%d pages swap cached\n", cached);
519 printk("Total of %ld pages in page table cache\n", pgtable_cache_size);
520 printk("%d free buffer pages\n", nr_free_buffer_pages());
524 * call_pernode_memory - use SRAT to call callback functions with node info
525 * @start: physical start of range
526 * @len: length of range
527 * @arg: function to call for each range
529 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
530 * out to which node a block of memory belongs. Ignore memory that we cannot
531 * identify, and split blocks that run across multiple nodes.
533 * Take this opportunity to round the start address up and the end address
534 * down to page boundaries.
536 void call_pernode_memory(unsigned long start, unsigned long len, void *arg)
538 unsigned long rs, re, end = start + len;
539 void (*func)(unsigned long, unsigned long, int);
542 start = PAGE_ALIGN(start);
549 if (!num_node_memblks) {
550 /* No SRAT table, so assume one node (node 0) */
552 (*func)(start, end - start, 0);
556 for (i = 0; i < num_node_memblks; i++) {
557 rs = max(start, node_memblk[i].start_paddr);
558 re = min(end, node_memblk[i].start_paddr +
559 node_memblk[i].size);
562 (*func)(rs, re - rs, node_memblk[i].nid);
570 * count_node_pages - callback to build per-node memory info structures
571 * @start: physical start of range
572 * @len: length of range
573 * @node: node where this range resides
575 * Each node has it's own number of physical pages, DMAable pages, start, and
576 * end page frame number. This routine will be called by call_pernode_memory()
577 * for each piece of usable memory and will setup these values for each node.
578 * Very similar to build_maps().
580 static int count_node_pages(unsigned long start, unsigned long len, int node)
582 unsigned long end = start + len;
584 mem_data[node].num_physpages += len >> PAGE_SHIFT;
585 if (start <= __pa(MAX_DMA_ADDRESS))
586 mem_data[node].num_dma_physpages +=
587 (min(end, __pa(MAX_DMA_ADDRESS)) - start) >>PAGE_SHIFT;
588 start = GRANULEROUNDDOWN(start);
589 start = ORDERROUNDDOWN(start);
590 end = GRANULEROUNDUP(end);
591 mem_data[node].max_pfn = max(mem_data[node].max_pfn,
593 mem_data[node].min_pfn = min(mem_data[node].min_pfn,
594 start >> PAGE_SHIFT);
600 * paging_init - setup page tables
602 * paging_init() sets up the page tables for each node of the system and frees
603 * the bootmem allocator memory for general use.
605 void paging_init(void)
607 unsigned long max_dma;
608 unsigned long zones_size[MAX_NR_ZONES];
609 unsigned long zholes_size[MAX_NR_ZONES];
610 unsigned long max_gap, pfn_offset = 0;
613 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
615 efi_memmap_walk(find_largest_hole, &max_gap);
617 /* so min() will work in count_node_pages */
618 for (node = 0; node < numnodes; node++)
619 mem_data[node].min_pfn = ~0UL;
621 efi_memmap_walk(filter_rsvd_memory, count_node_pages);
623 for (node = 0; node < numnodes; node++) {
624 memset(zones_size, 0, sizeof(zones_size));
625 memset(zholes_size, 0, sizeof(zholes_size));
627 num_physpages += mem_data[node].num_physpages;
629 if (mem_data[node].min_pfn >= max_dma) {
630 /* All of this node's memory is above ZONE_DMA */
631 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
632 mem_data[node].min_pfn;
633 zholes_size[ZONE_NORMAL] = mem_data[node].max_pfn -
634 mem_data[node].min_pfn -
635 mem_data[node].num_physpages;
636 } else if (mem_data[node].max_pfn < max_dma) {
637 /* All of this node's memory is in ZONE_DMA */
638 zones_size[ZONE_DMA] = mem_data[node].max_pfn -
639 mem_data[node].min_pfn;
640 zholes_size[ZONE_DMA] = mem_data[node].max_pfn -
641 mem_data[node].min_pfn -
642 mem_data[node].num_dma_physpages;
644 /* This node has memory in both zones */
645 zones_size[ZONE_DMA] = max_dma -
646 mem_data[node].min_pfn;
647 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
648 mem_data[node].num_dma_physpages;
649 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
651 zholes_size[ZONE_NORMAL] = zones_size[ZONE_NORMAL] -
652 (mem_data[node].num_physpages -
653 mem_data[node].num_dma_physpages);
658 PAGE_ALIGN(max_low_pfn * sizeof(struct page));
659 vmem_map = (struct page *) vmalloc_end;
661 efi_memmap_walk(create_mem_map_page_table, 0);
662 printk("Virtual mem_map starts at 0x%p\n", vmem_map);
665 pfn_offset = mem_data[node].min_pfn;
667 free_area_init_node(node, NODE_DATA(node),
668 vmem_map + pfn_offset, zones_size,
669 pfn_offset, zholes_size);
672 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page));