discontig.c

来自「Linux Kernel 2.6.9 for OMAP1710」· C语言 代码 · 共 696 行 · 第 1/2 页

/*
 * Copyright (c) 2000, 2003 Silicon Graphics, Inc.  All rights reserved.
 * Copyright (c) 2001 Intel Corp.
 * Copyright (c) 2001 Tony Luck <[e-mail]>
 * Copyright (c) 2002 NEC Corp.
 * Copyright (c) 2002 Kimio Suganuma <[e-mail]>
 */

/*
 * Platform initialization for Discontig Memory
 */

#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/swap.h>
#include <linux/bootmem.h>
#include <linux/acpi.h>
#include <linux/efi.h>
#include <asm/pgalloc.h>
#include <asm/tlb.h>
#include <asm/meminit.h>
#include <asm/numa.h>
#include <asm/sections.h>

/*
 * Track per-node information needed to setup the boot memory allocator, the
 * per-node areas, and the real VM.
 */
struct early_node_data {
	struct ia64_node_data *node_data;
	pg_data_t *pgdat;
	unsigned long pernode_addr;
	unsigned long pernode_size;
	struct bootmem_data bootmem_data;
	unsigned long num_physpages;
	unsigned long num_dma_physpages;
	unsigned long min_pfn;
	unsigned long max_pfn;
};

static struct early_node_data mem_data[NR_NODES] __initdata;

/**
 * reassign_cpu_only_nodes - called from find_memory to move CPU-only nodes to a memory node
 *
 * This function will move nodes with only CPUs (no memory)
 * to a node with memory which is at the minimum numa_slit distance.
 * Any reassigments will result in the compression of the nodes
 * and renumbering the nid values where appropriate.
 * The static declarations below are to avoid large stack size which
 * makes the code not re-entrant.
 */
static void __init reassign_cpu_only_nodes(void)
{
	struct node_memblk_s *p;
	int i, j, k, nnode, nid, cpu, cpunid, pxm;
	u8 cslit, slit;
	static DECLARE_BITMAP(nodes_with_mem, NR_NODES) __initdata;
	static u8 numa_slit_fix[MAX_NUMNODES * MAX_NUMNODES] __initdata;
	static int node_flip[NR_NODES] __initdata;
	static int old_nid_map[NR_CPUS] __initdata;

	for (nnode = 0, p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
		if (!test_bit(p->nid, (void *) nodes_with_mem)) {
			set_bit(p->nid, (void *) nodes_with_mem);
			nnode++;
		}

	/*
	 * All nids with memory.
	 */
	if (nnode == numnodes)
		return;

	/*
	 * Change nids and attempt to migrate CPU-only nodes
	 * to the best numa_slit (closest neighbor) possible.
	 * For reassigned CPU nodes a nid can't be arrived at
	 * until after this loop because the target nid's new
	 * identity might not have been established yet. So
	 * new nid values are fabricated above numnodes and
	 * mapped back later to their true value.
	 */
	for (nid = 0, i = 0; i < numnodes; i++)  {
		if (test_bit(i, (void *) nodes_with_mem)) {
			/*
			 * Save original nid value for numa_slit
			 * fixup and node_cpuid reassignments.
			 */
			node_flip[nid] = i;

			if (i == nid) {
				nid++;
				continue;
			}

			for (p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
				if (p->nid == i)
					p->nid = nid;

			cpunid = nid;
			nid++;
		} else
			cpunid = numnodes;

		for (cpu = 0; cpu < NR_CPUS; cpu++)
			if (node_cpuid[cpu].nid == i) {
				/*
				 * For nodes not being reassigned just
				 * fix the cpu's nid and reverse pxm map
				 */
				if (cpunid < numnodes) {
					pxm = nid_to_pxm_map[i];
					pxm_to_nid_map[pxm] =
					          node_cpuid[cpu].nid = cpunid;
					continue;
				}

				/*
				 * For nodes being reassigned, find best node by
				 * numa_slit information and then make a temporary
				 * nid value based on current nid and numnodes.
				 */
				for (slit = 0xff, k = numnodes + numnodes, j = 0; j < numnodes; j++)
					if (i == j)
						continue;
					else if (test_bit(j, (void *) nodes_with_mem)) {
						cslit = numa_slit[i * numnodes + j];
						if (cslit < slit) {
							k = numnodes + j;
							slit = cslit;
						}
					}

				/* save old nid map so we can update the pxm */
				old_nid_map[cpu] = node_cpuid[cpu].nid;
				node_cpuid[cpu].nid = k;
			}
	}

	/*
	 * Fixup temporary nid values for CPU-only nodes.
	 */
	for (cpu = 0; cpu < NR_CPUS; cpu++)
		if (node_cpuid[cpu].nid == (numnodes + numnodes)) {
			pxm = nid_to_pxm_map[old_nid_map[cpu]];
			pxm_to_nid_map[pxm] = node_cpuid[cpu].nid = nnode - 1;
		} else {
			for (i = 0; i < nnode; i++) {
				if (node_flip[i] != (node_cpuid[cpu].nid - numnodes))
					continue;

				pxm = nid_to_pxm_map[old_nid_map[cpu]];
				pxm_to_nid_map[pxm] = node_cpuid[cpu].nid = i;
				break;
			}
		}

	/*
	 * Fix numa_slit by compressing from larger
	 * nid array to reduced nid array.
	 */
	for (i = 0; i < nnode; i++)
		for (j = 0; j < nnode; j++)
			numa_slit_fix[i * nnode + j] =
				numa_slit[node_flip[i] * numnodes + node_flip[j]];

	memcpy(numa_slit, numa_slit_fix, sizeof (numa_slit));

	for (i = nnode; i < numnodes; i++)
		node_set_offline(i);

	numnodes = nnode;

	return;
}

/*
 * To prevent cache aliasing effects, align per-node structures so that they
 * start at addresses that are strided by node number.
 */
#define NODEDATA_ALIGN(addr, node)						\
	((((addr) + 1024*1024-1) & ~(1024*1024-1)) + (node)*PERCPU_PAGE_SIZE)

/**
 * build_node_maps - callback to setup bootmem structs for each node
 * @start: physical start of range
 * @len: length of range
 * @node: node where this range resides
 *
 * We allocate a struct bootmem_data for each piece of memory that we wish to
 * treat as a virtually contiguous block (i.e. each node). Each such block
 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
 * if necessary.  Any non-existent pages will simply be part of the virtual
 * memmap.  We also update min_low_pfn and max_low_pfn here as we receive
 * memory ranges from the caller.
 */
static int __init build_node_maps(unsigned long start, unsigned long len,
				  int node)
{
	unsigned long cstart, epfn, end = start + len;
	struct bootmem_data *bdp = &mem_data[node].bootmem_data;

	epfn = GRANULEROUNDUP(end) >> PAGE_SHIFT;
	cstart = GRANULEROUNDDOWN(start);

	if (!bdp->node_low_pfn) {
		bdp->node_boot_start = cstart;
		bdp->node_low_pfn = epfn;
	} else {
		bdp->node_boot_start = min(cstart, bdp->node_boot_start);
		bdp->node_low_pfn = max(epfn, bdp->node_low_pfn);
	}

	min_low_pfn = min(min_low_pfn, bdp->node_boot_start>>PAGE_SHIFT);
	max_low_pfn = max(max_low_pfn, bdp->node_low_pfn);

	return 0;
}

/**
 * early_nr_cpus_node - return number of cpus on a given node
 * @node: node to check
 *
 * Count the number of cpus on @node.  We can't use nr_cpus_node() yet because
 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
 * called yet.
 */
static int early_nr_cpus_node(int node)
{
	int cpu, n = 0;

	for (cpu = 0; cpu < NR_CPUS; cpu++)
		if (node == node_cpuid[cpu].nid)
			n++;

	return n;
}

/**
 * find_pernode_space - allocate memory for memory map and per-node structures
 * @start: physical start of range
 * @len: length of range
 * @node: node where this range resides
 *
 * This routine reserves space for the per-cpu data struct, the list of
 * pg_data_ts and the per-node data struct.  Each node will have something like
 * the following in the first chunk of addr. space large enough to hold it.
 *
 *    ________________________
 *   |                        |
 *   |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
 *   |    PERCPU_PAGE_SIZE *  |     start and length big enough
 *   |        NR_CPUS         |
 *   |------------------------|
 *   |   local pg_data_t *    |
 *   |------------------------|
 *   |  local ia64_node_data  |
 *   |------------------------|
 *   |          ???           |
 *   |________________________|
 *
 * Once this space has been set aside, the bootmem maps are initialized.  We
 * could probably move the allocation of the per-cpu and ia64_node_data space
 * outside of this function and use alloc_bootmem_node(), but doing it here
 * is straightforward and we get the alignments we want so...
 */
static int __init find_pernode_space(unsigned long start, unsigned long len,
				     int node)
{
	unsigned long epfn, cpu, cpus;
	unsigned long pernodesize = 0, pernode, pages, mapsize;
	void *cpu_data;
	struct bootmem_data *bdp = &mem_data[node].bootmem_data;

	epfn = (start + len) >> PAGE_SHIFT;

	pages = bdp->node_low_pfn - (bdp->node_boot_start >> PAGE_SHIFT);
	mapsize = bootmem_bootmap_pages(pages) << PAGE_SHIFT;

	/*
	 * Make sure this memory falls within this node's usable memory
	 * since we may have thrown some away in build_maps().
	 */
	if (start < bdp->node_boot_start || epfn > bdp->node_low_pfn)
		return 0;

	/* Don't setup this node's local space twice... */
	if (mem_data[node].pernode_addr)
		return 0;

	/*
	 * Calculate total size needed, incl. what's necessary
	 * for good alignment and alias prevention.
	 */
	cpus = early_nr_cpus_node(node);
	pernodesize += PERCPU_PAGE_SIZE * cpus;
	pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
	pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
	pernodesize = PAGE_ALIGN(pernodesize);
	pernode = NODEDATA_ALIGN(start, node);

	/* Is this range big enough for what we want to store here? */
	if (start + len > (pernode + pernodesize + mapsize)) {
		mem_data[node].pernode_addr = pernode;
		mem_data[node].pernode_size = pernodesize;
		memset(__va(pernode), 0, pernodesize);

		cpu_data = (void *)pernode;
		pernode += PERCPU_PAGE_SIZE * cpus;

		mem_data[node].pgdat = __va(pernode);
		pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));

		mem_data[node].node_data = __va(pernode);
		pernode += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));

		mem_data[node].pgdat->bdata = bdp;
		pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));

		/*
		 * Copy the static per-cpu data into the region we
		 * just set aside and then setup __per_cpu_offset
		 * for each CPU on this node.
		 */
		for (cpu = 0; cpu < NR_CPUS; cpu++) {
			if (node == node_cpuid[cpu].nid) {
				memcpy(__va(cpu_data), __phys_per_cpu_start,
				       __per_cpu_end - __per_cpu_start);
				__per_cpu_offset[cpu] = (char*)__va(cpu_data) -
					__per_cpu_start;
				cpu_data += PERCPU_PAGE_SIZE;
			}
		}
	}

	return 0;
}

/**
 * free_node_bootmem - free bootmem allocator memory for use
 * @start: physical start of range
 * @len: length of range
 * @node: node where this range resides
 *
 * Simply calls the bootmem allocator to free the specified ranged from
 * the given pg_data_t's bdata struct.  After this function has been called
 * for all the entries in the EFI memory map, the bootmem allocator will