path: root/arch/ia64/mm/discontig.c

/*
 * Copyright (c) 2000, 2003 Silicon Graphics, Inc.  All rights reserved.
 * Copyright (c) 2001 Intel Corp.
 * Copyright (c) 2001 Tony Luck <tony.luck@intel.com>
 * Copyright (c) 2002 NEC Corp.
 * Copyright (c) 2002 Kimio Suganuma <k-suganuma@da.jp.nec.com>
 * Copyright (c) 2004 Silicon Graphics, Inc
 *	Russ Anderson <rja@sgi.com>
 *	Jesse Barnes <jbarnes@sgi.com>
 *	Jack Steiner <steiner@sgi.com>
 */

/*
 * 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;

/**
 * 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, MAX_NUMNODES) __initdata;
	static u8 numa_slit_fix[MAX_NUMNODES * MAX_NUMNODES] __initdata;
	static int node_flip[MAX_NUMNODES] __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 == num_online_nodes())
		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 num_online_nodes() and
	 * mapped back later to their true value.
	 */
	/* MCD - This code is a bit complicated, but may be unnecessary now.
	 * We can now handle much more interesting node-numbering.
	 * The old requirement that 0 <= nid <= numnodes <= MAX_NUMNODES
	 * and that there be no holes in the numbering 0..numnodes
	 * has become simply 0 <= nid <= MAX_NUMNODES.
	 */
	nid = 0;
	for_each_online_node(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 = MAX_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 < MAX_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 num_online_nodes().
				 */
				slit = 0xff;
				k = 2*num_online_nodes();
				for_each_online_node(j) {
					if (i == j)
						continue;
					else if (test_bit(j, (void *) nodes_with_mem)) {
						cslit = numa_slit[i * num_online_nodes() + j];
						if (cslit < slit) {
							k = num_online_nodes() + 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 == (2*num_online_nodes())) {
			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 - num_online_nodes()))
					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] * num_online_nodes() + node_flip[j]];

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

	nodes_clear(node_online_map);
	for (i = 0; i < nnode; i++)
		node_set_online(i);

	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