discontig.c

linux-2.6.15.6

/*
 * 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]>
 * Copyright (c) 2004 Silicon Graphics, Inc
 *	Russ Anderson <[e-mail]>
 *	Jesse Barnes <[e-mail]>
 *	Jack Steiner <[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 <linux/nodemask.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[MAX_NUMNODES] __initdata;
static nodemask_t memory_less_mask __initdata;

/*
 * To prevent cache aliasing effects, align per-node structures so that they
 * start at addresses that are strided by node number.
 */
#define MAX_NODE_ALIGN_OFFSET	(32 * 1024 * 1024)
#define NODEDATA_ALIGN(addr, node)						\
	((((addr) + 1024*1024-1) & ~(1024*1024-1)) + 				\
	     (((node)*PERCPU_PAGE_SIZE) & (MAX_NODE_ALIGN_OFFSET - 1)))

/**
 * 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.  Note that node 0 will also count all non-existent cpus.
 */
static int __init 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;
}

/**
 * compute_pernodesize - compute size of pernode data
 * @node: the node id.
 */
static unsigned long __init compute_pernodesize(int node)
{
	unsigned long pernodesize = 0, cpus;

	cpus = early_nr_cpus_node(node);
	pernodesize += PERCPU_PAGE_SIZE * cpus;
	pernodesize += node * L1_CACHE_BYTES;
	pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
	pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
	pernodesize = PAGE_ALIGN(pernodesize);
	return pernodesize;
}

/**
 * per_cpu_node_setup - setup per-cpu areas on each node
 * @cpu_data: per-cpu area on this node
 * @node: node to setup
 *
 * 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.  Return a pointer to
 * the end of the area.
 */
static void *per_cpu_node_setup(void *cpu_data, int node)
{
#ifdef CONFIG_SMP
	int cpu;

	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;
		}
	}
#endif
	return cpu_data;
}

/**
 * fill_pernode - initialize pernode data.
 * @node: the node id.
 * @pernode: physical address of pernode data
 * @pernodesize: size of the pernode data
 */
static void __init fill_pernode(int node, unsigned long pernode,
	unsigned long pernodesize)
{
	void *cpu_data;
	int cpus = early_nr_cpus_node(node);
	struct bootmem_data *bdp = &mem_data[node].bootmem_data;

	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;
	pernode += node * L1_CACHE_BYTES;

	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));

	cpu_data = per_cpu_node_setup(cpu_data, node);

	return;
}

/**
 * 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
 *   |    cpus_on_this_node   | Node 0 will also have entries for all non-existent 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;
	unsigned long pernodesize = 0, pernode, pages, mapsize;
	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.
	 */
	pernodesize = compute_pernodesize(node);
	pernode = NODEDATA_ALIGN(start, node);

	/* Is this range big enough for what we want to store here? */
	if (start + len > (pernode + pernodesize + mapsize))
		fill_pernode(node, pernode, pernodesize);

	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
 * be ready to service allocation requests.
 */
static int __init free_node_bootmem(unsigned long start, unsigned long len,
				    int node)
{
	free_bootmem_node(mem_data[node].pgdat, start, len);

	return 0;
}

/**
 * reserve_pernode_space - reserve memory for per-node space
 *
 * Reserve the space used by the bootmem maps & per-node space in the boot
 * allocator so that when we actually create the real mem maps we don't
 * use their memory.
 */
static void __init reserve_pernode_space(void)
{
	unsigned long base, size, pages;
	struct bootmem_data *bdp;
	int node;

	for_each_online_node(node) {
		pg_data_t *pdp = mem_data[node].pgdat;

		if (node_isset(node, memory_less_mask))
			continue;

		bdp = pdp->bdata;

		/* First the bootmem_map itself */
		pages = bdp->node_low_pfn - (bdp->node_boot_start>>PAGE_SHIFT);
		size = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
		base = __pa(bdp->node_bootmem_map);
		reserve_bootmem_node(pdp, base, size);

		/* Now the per-node space */
		size = mem_data[node].pernode_size;
		base = __pa(mem_data[node].pernode_addr);
		reserve_bootmem_node(pdp, base, size);
	}
}

/**
 * initialize_pernode_data - fixup per-cpu & per-node pointers
 *
 * Each node's per-node area has a copy of the global pg_data_t list, so
 * we copy that to each node here, as well as setting the per-cpu pointer
 * to the local node data structure.  The active_cpus field of the per-node
 * structure gets setup by the platform_cpu_init() function later.
 */
static void __init initialize_pernode_data(void)
{
	pg_data_t *pgdat_list[MAX_NUMNODES];
	int cpu, node;

	for_each_online_node(node)
		pgdat_list[node] = mem_data[node].pgdat;

	/* Copy the pg_data_t list to each node and init the node field */
	for_each_online_node(node) {
		memcpy(mem_data[node].node_data->pg_data_ptrs, pgdat_list,
		       sizeof(pgdat_list));
	}
#ifdef CONFIG_SMP
	/* Set the node_data pointer for each per-cpu struct */
	for (cpu = 0; cpu < NR_CPUS; cpu++) {
		node = node_cpuid[cpu].nid;
		per_cpu(cpu_info, cpu).node_data = mem_data[node].node_data;
	}
#else
	{
		struct cpuinfo_ia64 *cpu0_cpu_info;
		cpu = 0;
		node = node_cpuid[cpu].nid;
		cpu0_cpu_info = (struct cpuinfo_ia64 *)(__phys_per_cpu_start +
			((char *)&per_cpu__cpu_info - __per_cpu_start));
		cpu0_cpu_info->node_data = mem_data[node].node_data;
	}
#endif /* CONFIG_SMP */
}

/**
 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit
 * 	node but fall back to any other node when __alloc_bootmem_node fails
 *	for best.
 * @nid: node id
 * @pernodesize: size of this node's pernode data
 */
static void __init *memory_less_node_alloc(int nid, unsigned long pernodesize)
{
	void *ptr = NULL;
	u8 best = 0xff;
	int bestnode = -1, node, anynode = 0;

	for_each_online_node(node) {
		if (node_isset(node, memory_less_mask))
			continue;
		else if (node_distance(nid, node) < best) {
			best = node_distance(nid, node);
			bestnode = node;
		}
		anynode = node;
	}

	if (bestnode == -1)
		bestnode = anynode;

	ptr = __alloc_bootmem_node(mem_data[bestnode].pgdat, pernodesize,
		PERCPU_PAGE_SIZE, __pa(MAX_DMA_ADDRESS));